We are thrilled to announce our recent lead investment in Heymax.ai’s US$2.6 million Seed financing round alongside Tenity and key strategic angels. Heymax.ai is an ambitious platform that aims to put consumers first, and deliver value to merchants and loyalty reward operators. Heymax.ai’s team today is focused on helping consumers generate more value across everyday spending — long term, their mission is far larger.

We look forward to partnering with Joe Lu, Wang Ke, Jialu Zhong and Sean Dy — in their ambitions to build an interoperable global currency for loyalty and help consumers further optimize and aggregate reward from their spending, to earn a free trip every year.

Market Opportunity & Thesis

The continued untapped opportunity in optimizing consumer finance.

Today, more than 75% of consumers make concerted efforts with regards to their personal finance management (“PFM”) (i.e. making regular budgets, developing credit card optimization strategies, engaging in cashback programs) but less than 10% report finding it easy. The key issues reported generally relate to the fragmentation of data with regards to one’s personal transactional flow (across various cards, bank accounts, and billing systems), making it difficult to consolidate their spending, let alone optimize said spend.

Since the advent of the internet (and subsequently, mobile) there have been multiple attempts by emerging technology firms to tackle PFM — many of these platforms focused on an individual’s personal “P&L” from the outset, which required substantive data input and tracking. For the most part, these platforms have not scaled meaningfully and struggle with capturing long term value for consumers.

This is despite the market opportunity for PFM being immense — over the past decade, we have spent time reviewing several PFM solutions but found that the value creation for the average consumer lacking. It is not enough for an application to tell you how much you spent as a “wedge” — rather it is important to show value (i.e., cashback, savings, miles) from day one.

The ongoing search for localized merchant demand generation.

Over the past decade, we have also observed the channels available to merchants to attract consumers get narrower — today, Meta and Google’s digital advertising combined market share exceeds 50.0% of the entire market (with the rest covered by the likes of TikTok and Apple). This has pushed up customer acquisition costs and lowered the efficiency of marketing dollars.

In today’s current market environment (where most businesses are tightening spend), merchants are seeking innovative ways to attract and retain customers. At January Capital, we have long believed that there is a need for localized demand generation solutions in the Asia-Pacific region — we have been fortunate to invest and work with Henry and Joel at Shopback since 2017. Shopback has done an incredible job of building market leadership in the ‘value shopper’ oriented segment and today dominates the cashback category across the Asia-Pacific region.

But there remains work to be done — there are other segments of shoppers whom merchants are eager to engage with and will pay for conversions that yield from traffic sources. In this vein, the value perceived by consumers between cashback and miles as a form of loyalty rewards is very different. Notably, the former incentivizes spend for immediate return, while the latter is a value accruement over time targeted at a larger expenditure later (usually travel and experiences). Henceforth, by extension, the nature of a longer-termed loyalty rewards program naturally provides a strong “wedge” into helping consumers manage personal finances through optimizing their spend over a longer and more sustained period of time.

We maintain that the more localized demand generation platforms there are in the ecosystem, the better it is for both consumers and merchants as it provides more options for alternative channel versus current acquisition channels (which has diminishing returns over time). In a previous Investment Note for Twid (see here) in 2022, we noted an increasing fragmentation of reward points with established ecommerce brands and platform building their own loyalty programs; this has only continue to accelerate.

Reimagining reward programs through interoperability

Globally, companies worldwide have been increasing their slice of loyalty marketing budget to continuously retain and attract new customers. Reward programs in and of itself is not a new initiative. Starbucks Rewards has been phenomenally successfully in incentivizing consumers to spend within its ecosystem. Closer to home, Shopee Coins has creatively engaged consumers to accrue further rewards through spending. Still, these are all rewards that are locked within its closed ecosystem and merchants or brands without such scale will not materially drive down customer acquisition cost.

At January Capital, we’ve consistently had a view on a more ‘open’ approach in allowing consumers to monetize their rewards. It is likely that major platforms and brands will only continue to ring-fence their ecosystem through closed-loop reward programs. Such continued hyper-fragmentation of loyalty programs will only further commoditize reward points, drive up cost of customer acquisition for merchants, and increase the mental load on consumers who are seeking to optimize their spending power.

Here, Heymax.ai creatively solves for ‘openness’ through its ambitious vision of making Max Miles the premium loyalty currency (earn) for consumers to accrue easily, and subsequently allowing consumers to flexibility redeem for any flights or convert into miles or other loyalty points (burn). Heymax.ai is in effect a three-sided marketplace that:

1. provides an aggregated platform for consumers to accrue loyalty points,

2. an effective channel for customer acquisition versus lower ROI marketing initiatives and

3. leveraging first-party data to help loyalty operators retain high value customers.

And they achieve all that by winning consumers first.

Connecting consumers and brands through a common loyalty currency

Today, consumers who want to optimize and maximize their rewards are flooded with a flurry of choices. The cognitive load to min-max miles accrual by pairing credit card transactions to MCC codes means that points optimization is usually limited to a handful of hardcore consumers. Consumers in this camp tend to juggle multiple credit cards (upwards of 10+) and have built complex Excel trackers to monitor redemption cycles.

The Heymax.ai team has a nuanced understanding of building relevant tools to aid consumers in their loyalty points earn and burn endeavor. This means that Heymax.ai has been able to attract an initial wave of points junkie who are fervent users of the various tools Heymax.ai has built, but also bring others along on the miles accrual journey.

Heymax.ai does this by simplifying the entire earn and burn process into a simple mechanism:

1. Easily understood reward system: Building interoperability across multiple loyalty programs allows them to abstract disparate reward program currency into their own bespoke currency: Max Miles.

2. Aggregating the “Earn” of rewards: Heymax.ai grants consumers access to more than 500+ major merchants across key categories like travel, retail and F&B, available via their multi-channel platform. Heymax.ai have also successfully built distribution channels via partnerships with Singsaver and Airwallex to expand consumers’ MaxMiles earning channels.

3. Layer of Optimization: In a partnership with Visa, Heymax.ai helps consumers figure out the best credit card to use for each transaction to maximize their existing credit card rewards. They also simplify the management of multiple cards for consumers and streamlines rewards tracking.

4. Smoothening “Burn” of rewards: for too long, redemption process from existing credit card programs have been clunky and untransparent. With Heymax.ai’s newly launched FlyAnywhere feature, consumers can redeem any flight with a best-in-class miles transfer value — becoming a true open-loop outcome for consumers.

Brands and merchants have for a long time been trying to build their own loyalty programs. There is a long tail graveyard of programs that have not achieved their purpose. We feel that there is space for an independent provider to build a layer of interoperability between loyalty programs and allow users to orchestrate their ‘burn’ and aggregate their ‘earn’ through Heymax.ai’s platform. Heymax.ai is early in thinking about how to generate value in a tripartite way, but there are early signs of them being able to prove out how they can deliver such value to consumers and brands at scale.

Although Heymax.ai is still early in its journey, it has already won over a core group of consumers who are strong advocates of the Heymax.ai platform. January Capital believes that Heymax.ai has a strong right to become the dominant loyalty currency of the world through its nuanced efficacy in smoothening out the Earn & Burn process.

Product Overview & Vision

Heymax.ai’s commerce platform enables consumers to “double-dip” on loyalty points through accruing Max Miles and optimizing existing credit card rewards. Consumers can easily sign up for a Heymax.ai account through the webapp, mobile and install a simple chrome extension that alerts consumers to check out via Heymax.ai to earn additional rewards.

When consumers “Shop with Heymax.ai”, they are directed to Heymax.ai’s integrated partner merchants to make purchases without any manual heavy lifting. Upon a successful purchase, consumers are rewarded with Max Miles that has no expiry and is transferable to airlines or hotels’ loyalty programs at a 1:1 transfer ratio. Furthermore, consumers can tag their existing credit cards on the platform to get instant recommendation on the best credit card to use for that purchase to maximize their rewards. Beyond that, consumers can purchase vouchers from their favorite merchants at favorable rates instead of making an unrewarded purchase.

The Team

January Capital has a long-held thesis on “localized demand generation” in commerce and “decomplexifying the financial stack for broad access” in our core geography focus. Heymax.ai’s long-term ambition sits in the cross-intersection of two of our key pillars. The complexity of building a global platform with a foundational blend between commerce and fintech means that we need to look for founders who we believe not only embody their vision but are also excellent operators who can execute at high finesse.

From the moment we met Heymax.ai’s founding team, we were enthralled by the teams’ imaginative goal and their combined experience operating at large technology companies. Heymax.ai is led by Joe Lu, Wang Ke, Jialu Zhong and Sean Dy, who have worked together for many years at Meta in various engineering functions in San Francisco before moving to the Meta regional HQ in Singapore.

Joe is one of the rare founders who is building a company in the mold of what he himself would want to use as a user. He is at the forefront of miles maximization as a consumer. Since the age of 18, Joe has earned and burned more than 15.0 million miles, accumulated more than SGD $5000 of cashback on Shopback (a January Capital portfolio company) and uses the most obscure open-source tools to maximize all his various credit card rewards. He can intimately “speak the language” with Heymax.ai’s core users who are also fervent miles junkies. Beyond his personal obsession, Joe has a decade of engineering leadership experience at Meta and Expedia Group.

Wang Ke, meanwhile, holds a PhD from National University of Singapore with his thesis centered on studying applications of utilizing reinforcement learning in mobile networks. Post academia, he held positions as a senior software engineer at Quantcast, Indeed and most recently Meta.

Jialu Zhong, has been part of building some of the most innovative products in the world. She graduated with a master’s degree in computer science from Columbia University. Thereafter, she joined Uber in San Francisco before being approached to take up a role at Meta with the team at Instagram.

Sean Dy, is a close to 10 years veteran at Meta having joined them immediately after graduation as a Data Engineer. He served across Meta’s San Francisco, New York and Singapore offices.

Progress so far and journey ahead

Today, Heymax.ai has more than 500 merchants partners, 20+ loyalty partners and has grown beyond 50,000 users within a few months of launch. They have already expanded into and are building up a presence in Australia with a path to launch in other markets quickly. Since launch, Heymax.ai has been building an active and rapidly growing community — join the more than 5000 Telegram members to learn more.

While still early in the journey, Heymax.ai has considerable organic growth borne from driving a wedge into a segment (miles optimization) that has long been underserved with lackluster Earn and Burn mechanisms. We are looking forward to adopting Max Miles as a loyalty currency to enable our team to travel on.

We are excited about what the future holds for Heymax.ai, along with their key strategic angel investors.

For more information about January Capital, please visit our website.

For more information about Heymax, please visit their website.

DISCLAIMER

January Capital is a licensed fund manager in both Singapore (through Jan Cap Pte. Ltd.) and Australia (through January Capital Pty Ltd) (together, the “Managers”) licensed under the purview of the Monetary Authority of Singapore and the Australian Securities and Investments Commission respectively.

The information contained in this article is being provided to you for informational purposes only and does not constitute in any circumstance an offer to sell, or a solicitation of an offer to buy any interest in the funds managed by the Managers or any other companies mentioned herein. Additionally, the information in this article and any and all forward-looking statements are based upon assumptions that may not prove to be correct or accurate and the actual financial results/performance and opportunities (amongst others) may differ materially from these statements. As such, nothing contained in this article should be construed as or relied upon as financial, legal or taxation advice, or an encouragement to invest in our funds or any companies mentioned in this article. Past performance is not indicative of future results.

We are thrilled to announce our investment in Sleek EV’s Pre-Series A financing round alongside Orzon Ventures (a venture capital fund led by OR and 500 TukTuks) and A2D Ventures. Sleek EV is a brand manufacturer of electric two-wheelers, with Thailand as its first target market. Sleek EV represents January Capital’s first investment in electric mobility infrastructure and we are excited to support Kantinan (Ben) Tunveenukoon and Zhang Quan (ZQ) Ong on their journey to accelerate Southeast Asia’s transition to electric two-wheelers.

The potential of the Southeast Asian electric two-wheeler (“e2W”) market

Southeast Asia’s congested streets in densely populated urban centres have compelled many consumers to seek out a mobility solution capable of efficiently navigating through relentless traffic. Along with economic constraints, consumers in these countries sought vehicles that are less expensive, require minimal maintenance and offer performance efficiency. Against this backdrop, two-wheelers have emerged as the go to mobility solution for the majority of Southeast Asians, with over 80% of households owning two-wheelers.

The potential of electric two-wheelers (“e2W”) have previously been hailed by some as a transformative force in the Southeast Asian two-wheeler industry by bringing benefits that include lower “fuel” costs, reduced maintenance requirements and lower carbon emissions, offering a better alternative to traditional internal combustion engine (“ICE”) counterparts. However, despite progress in the space, widespread adoption has historically encountered two main barriers — affordability and performance:

1. Affordability — historically, the total cost of ownership (“TCO”) for e2Ws surpassed that of ICE counterparts due to higher production costs, resulting in elevated purchase prices and higher taxes.

2. Performance — concerns regarding long-term performance, reliability, and convenience as compared to ICE two-wheelers due to the lack of supporting infrastructure, including charging/battery swapping stations and the accessibility of replacement parts.

While still in early stages of mainstream e2W adoption, Southeast Asia is starting to experience a transformative shift due to the widespread commercialization efforts that have lowered manufacturing costs and addressed affordability concerns. As a result, e2Ws have reached a pivotal juncture where pricing, which is a key factor driving decision making for most Southeast Asian consumers, is now competitive with that of traditional ICE two-wheelers.

Moreover, growing investments into charging stations, battery swapping technology and various orchestration software from strategic and financial investors alike have resulted in the expansion of infrastructure that supports the day-to-day use of e2Ws, whether for the individual vehicle owner, or businesses and governments looking to electrify its fleets.

Looking across the e2W value chain today, we believe the opportunity for investment primarily lies in those who are who are able to capture market share swiftly and effectively through their strong brand value, established distribution networks and superior customer service. Secondarily, there is potential in those engaged in building the necessary infrastructure to facilitate this transformation which include battery swapping/charging stations and software development.

The opportunity in Thailand’s two-wheeler market

From a geographic lens, the markets of Indonesia, Vietnam, Philippines, and Thailand represent large opportunities given how deeply ingrained two-wheelers are in each respective economy. With regards to e2Ws, Indonesian and Vietnamese markets are more progressed, funded and competed, with multiple notable startups, and even incumbents rolling out their bikes on the streets. In the Philippines, government initiatives, another key driver of adoption, has yet to catch up with the progress in other nations. This casts a light upon Thailand as a country where market timing aligns with the growing ambitions of new startups.

In line with the broader transportation landscape in Southeast Asia, two-wheelers serve as the primary mode of transportation in Thailand, with 21.7 million two-wheelers on the road, accounting for 51% of the total vehicles in circulation. In 2022 alone, 1.8 million new two-wheelers were sold in Thailand. While the majority of these were ICE two-wheelers, the proportion of e2Ws relative to prior periods grew substantially with a 150% Y-o-Y increase in e2W registrations.

More notable is the surge in newly registered e2Ws in Q1 CY2023 alone, where there was a remarkable 254% Y-o-Y growth. Today, e2Ws account for ~1% of all newly registered two-wheelers in Thailand, a material uplift from 0.3% market share just 12 months ago[1]. We believe the accelerating trend in the adoption of e2Ws in Thailand is being driven by:

1. Reducing disparity in total cost of ownership (“TCO”) — Although e2Ws upfront costs remain slightly more expensive than ICE two-wheelers, cost savings throughout the lifetime of the vehicle achieved through charging/battery swapping instead of fuel, and lower taxation expenses have resulted in cost parity, or in certain cases, even lower TCO for e2Ws.

2. Robust government initiatives — Thailand possesses the most holistic set of government incentives and subsidies in Southeast Asia, spanning across demand, supply, and infrastructure stimulators, which further encourages EV adoption.

3. Growing consumer awareness — Within Southeast Asia, Thai consumers exhibit a strong inclination towards adopting more fuel-efficient vehicles, trailing only behind Singapore, as 43% of consumers express their intentions to transition to such vehicles. Furthermore, Thai consumers display the highest level of interest in alternative powertrains for their future vehicles, with 14% favouring Battery Electric Vehicles (“BEVs”), making Thailand the leader in this preference among Southeast Asian countries.

With favourable tailwinds and no clear winner, Thailand’s e2W market represents an untapped opportunity waiting to be seized.

Electrifying Southeast Asia’s two-wheeler market with Sleek EV

Given the above opportunity, the founders of Sleek EV have chosen Thailand as their launch market to deliver cost-effective and premium performing e2Ws through an extensive distribution network, supported by partnerships across the ecosystem.

Sleek EV has embraced the completely knocked-down method as a supply chain solution, enabling Sleek EV to manufacture and distribute their e2Ws without having to invest significant capex from day one. This translates to competitive prices even through a dealership model, with the potential to expand directly into a D2C model in the future. In addition, Sleek EV’s partnerships with various financial institutions enables them to offer 0% downpayment to consumers (a conventional practice with ICE two-wheelers), which further empowers consumers to purchase from them.

Sleek EV also recognizes the value of building a strong distribution network and are continually investing significant resources to build and expand their distribution channels. These ongoing partnerships with premier local two-wheeler dealerships and enterprise clients have culminated in the sale of over 1,000 units (in the past 12 months). More importantly, having their e2Ws displayed in prominent dealerships across the country while being used by some of Thailand’s largest enterprises, has enabled Sleek EV to begin building strong brand equity which we believe will contribute significantly to building and solidifying their market position.

Lastly, viewing the two-wheeler ultimately as just a commodity, the team at Sleek EV understands the importance of building a moat which will allow them to remain competitive in the long run. The team believes that the way to build a moat is through a robust ecosystem accompanied with superior software compared to others available in the market today. Sleek EV’s ecosystem features e2W servicing centres offering the best-in-class after sales service, while leveraging partnerships with infrastructure providers to grant users access to a network of charging/battery-swapping stations. In terms of software, Sleek EV has built out their own Dealer Management System, Inventory Management system and App which links users directly with dealers — the outcome is a seamless user experience that benefits everyone in the Sleek ecosystem.

While Sleek EV is starting out as a brand manufacturer, we believe the opportunities lies in a future where consumer sentiment will shift towards viewing e2Ws not as an asset but as a commodity — one that can be easily replaced and financed. With this in mind, Sleek EV plans to leverage the data collected through its built-in operating system to identify opportunities around financing, insurance, and other after-sales services in order to offer superior experience that will retain customers through their lifetime.

Experienced co-founders operating in a familiar space

Ben and ZQ possess extensive experience in the EV industry; Ben was a founding member of Swag EV, the first e2W brand in Thailand, and was credited with starting their Thailand operations, scaling the team, and establishing their distribution channels. Both co-founders then met and started GEKO, which became the best performing dealer for the largest selling EV brand in Thailand, Deco Green Energy, accounting for ~15% of their total sales in the country. This is on top of being a fintech startup focused on financing e2Ws. Over a period of three years in GEKO, the team was able to set up operations, and through their Internet-of-Things (“IoT”) devices built into bikes, leverage data to offer e-KYC, AI-driven credit scoring, and flexible financing options for hundreds of riders.

Seeing the progress in e2W technology and the growing opportunity in Thailand, Ben and ZQ decided to take the next step, tapping into their understanding of the EV supply chain and financing to start Sleek EV. To date, Sleek EV’s management team has built a growing network of dealerships and strategic partnerships with financial institutions and ecosystem players, all while delivering strong top-line performance.

We at January Capital believe that lessons learnt, perspectives acquired, along with demonstrated execution capabilities uniquely positions the experienced Sleek EV management team to win out in this market.

Progress to date and the journey ahead

Since its launch in September 2022, Sleek EV has already sold over 1,000 e2Ws through their network of established dealerships and B2B clients across Thailand, boasting an impressive pipeline that could triple revenues by the end of 2023. With this round’s investment, Sleek EV is looking to supercharge its business by unlocking critical working capital and accelerate strategic hiring. More importantly, Sleek EV intends to unlock significant potential by leveraging their strategic investor’s ecosystem, granting them access to a broad customer base and extensive infrastructure support, positioning the company to reach new heights and solidify its position as a market leader.

We are excited about our partnership with Orzon Ventures and A2D Ventures as we join Sleek EV in its ambitions to electrify Thailand’s two-wheeler market and enable clean mobility throughout Southeast Asia.

For more information about January Capital, please visit our website and sign up for our regular newsletter.

For more information about Sleek EV please visit their website.

[1] January Capital Analysis.

At January Capital, we have been studying the implications of the proliferation of large language models — including legal, privacy and security, and efficacy, which we believe will result in a whole new wave of innovation accompanied by significant industry regulation. Today, we are open sourcing our internal work over the past year on this topic. This article is intended to be a summary of our emerging perspectives pertaining to the future of large language models, with the acknowledgement that many of the topics covered in this article are far from settled. If you are an industry practitioner or founder working in this space, we would love to hear from you.

You may reach out to our core team who is covering this space: Benjamin Dunphy (benjamin@january.capital), Simran Hathiramani (simran@january.capital) and Armaan Kapoor (armaan@january.capital).

A new technological era dawns

We are currently experiencing a moment in time that resembles significant shifts in technology, such as the emergence of the internet in 1994, the rise of mobile devices in 2007, and the introduction of cloud computing in 2010. Just as these technologies fundamentally transformed the way that people live, work, and play, we believe that artificial intelligence will revolutionize the way individuals and organizations operate.

In the last 6 months since ChatGPT was released, most of us have probably come across several content pieces about large language models (LLMs), such as technical papers, vendor landscapes, “how to” guides. AI continues to innovate at a ground-breaking pace and organizations (governments, enterprises, and startups) are racing to bolt on AI capabilities in their software stack as the time taken to launch AI features has never been shorter, due to few/zero shot learning. We are at an era where most knowledge work can be done declaratively (i.e. provide the outcome/end-state and AI can program and execute a series of steps to reach there), such as designing a website, conducting market research, or even generating an entire codebase. Just as software was eating the world a decade ago, AI is now consuming software.

While the potential of LLMs is undeniably vast, their imperfections cannot be overlooked. The primary reason behind this is the manner in which LLMs are trained — through unsupervised learning on massive datasets. As a result of the extensive world knowledge they acquire, users often encounter unpredictable outputs from a black box model and are susceptible to hallucinations — events where AI/LLMs confidently serve up responses that are wholly manufactured or inaccurate. Delving deeper into the profound effects of this paradigm shift reveals a multitude of concerns. For instance, the absence of a consensus regarding the regulation of intellectual property rights, and governing the use of proprietary company data into LLM interfaces like ChatGPT. Moreover, the malicious applications of AI, such as spear-phishing, mutable malware, and deepfakes, can be stimulated by these advanced language models. As LLM-powered applications and workflows become increasingly fundamental and pervasive in the knowledge workforce, it is paramount that AI systems are built and accessed in a safe, performant, and controllable manner.

Recent media coverage, such as Sam Altman (CEO of Open AI) warning against the potential pitfalls of leaving AI technology unregulated; as well as ongoing litigation such as the class action lawsuit against Microsoft, GitHub and OpenAI for allegedly violating copyright law by producing open-source code using AI — points towards a call for action and a need to shine a light on these potentially negative implications. In this regard, the UK has already committed £100m towards building capabilities for AI safety.

Quite possibly the number one item mentioned in daily technology news today is the rate at which LLMs have been adopted (compared to other technologies historically): as at date of writing, more than 100 million people are considered active users of ChatGPT. When the technology is adopted at such a rate (think Uber and ride-hailing regulations a decade ago), regulatory oversight and clarity on all risks understandably tend to lag. Today, there remains a lack of clarity on issues such as the terms of service, data privacy measures or nuanced infrastructure of the large language models powering these tools.

As we reflect on the past, it becomes evident that history often repeats itself. In this regard, we find it plausible to draw parallels between the concerns witnessed during the early stages of web, mobile, and cloud adoption. While this article does shed light on identified issues of LLM adoption to some extent, we firmly believe that they present significant opportunities for the emergence of category leaders. These opportunities manifest in four noteworthy emerging trends that we have observed:

1. Enterprises will prioritize enhancing the efficacy and precision of LLM-powered use cases. Control over the behaviour of LLMs will enable them to achieve the desired output and understand the reasons behind any deviations or hallucinations.
2. As employees rely on plugins and interfaces to harness the power of LLMs, there is a potential risk of transmitting proprietary data to platforms like ChatGPT. This introduces the concept of shadow AI, where sensitive information may unintentionally leak to third parties, exposing intellectual property and trade secrets.
3. The adoption of generative AI tools for adversarial purposes, such as phishing emails, deepfakes, and undetectable malware, poses an increased risk. These tools have the potential to deceive both humans and existing detection systems, thereby creating a new avenue for cyber attackers.
4. The lack of attribution to data and content creators or platforms that contribute to the training of LLMs. Prominent platforms like Reddit and StackOverflow, among others, play a significant role in shaping the knowledge and information within these language models, yet their contributions often go unacknowledged.

It’s important to highlight that some of these aforementioned challenges are not novel. Categories such as AI observability and data security/privacy have been in existence for the past few years as enterprises wrangle with AI and data. Nevertheless, the democratization of AI through LLMs has injected new life into the “why now” storyline. It has provided an exciting opportunity for these solutions to reshape themselves and cater to the demands of AI-native companies or enterprises seeking to embrace LLMs.

Emerging 2nd order effects and opportunities

Enhance the efficacy of LLMs

A key area of concern with enterprise LLM adoption is hallucinations, which occur when AI systems generate perceptual outputs that are not based on real-world input or are distorted representations of reality. An example of one such case was at the first demo of Google’s Bard chatbot, where the tool made false claim about the James Webb Space Telescope. In some instances, these hallucinations could have serious consequences. For instance, when two New York attorneys used ChatGPT to generate a legal motion which was filed in court, the LLM conjured up cases that did not exist. This can be largely attributable to a few reasons — the lack of reasoning by the language model, untrustworthy data that is fed into model training process (aka data poisoning), or insufficient fine tuning or control mechanisms.

For enterprises to successfully adopt and harness the value of these language models, they must “shepherd” the models into their desired use cases and behaviors from its vast knowledge and generic capabilities.

It is also worth noting that LLMs might possess societal bias characteristics, given the training data acquired from the internet can reflect gender, cultural, or historical biases. Dealing with AI bias is not new –we have witnessed cases including Apple’s credit card algorithm and Amazon’s automated resume screener in the pre-LLM era. While this is not a new problem, LLMs have the potential to perpetuate and amplify these biases.

In dealing with AI bias, regulators have typically fallen back on standard antidiscrimination legislation. Some enterprises have adopted AI explainability solutions to trace back on the root cause of AI bias. However, AI explainability is relatively easier when there are specific features within the model that could be attributed for generating the output or prediction. With large language models, providing insights on how the AI arrives at its decisions are significantly more complex given the vast number of parameters and the multiple layers within its neural network architecture.

Supervised fine tuning (“SFT”) and Reinforcement Learning with Human Feedback (“RLHF”)

Recently, there has been significant progress made in model training phases to steer pre-trained models towards generating the intended result and minimize hallucinations, known as supervised fine tuning (“SFT”) and reinforcement learning with human feedback (“RLHF”). Essentially, the pre-trained base model (e.g. GPT-3, Llama) is trained on custom datasets where humans evaluate and label the quality of the model’s output before feeding it back into the model to update its policy (by applying proximal policy optimization). New techniques to lower cost of fine tuning and improve quality of output are continuously unearthed by the industry/academia, such as Parameter Efficient Fine Tuning (PEFT) in the case of finetuning, and Direct Policy Optimization (DPO) in the case of simplifying RLHF pipelines.

Retrieval augmented generation (RAG)

Another common method of increasing output relevancy is to augment LLMs with additional public or private data sources. Imagine you are an E-Commerce brand and if you are building an LLM-powered customer support application, feeding in data sources like product catalogs, order information, and customer data would increase the relevancy of an AI-powered response. Following a prompt/query by the customer, RAG will look for relevant data across different sources which augments these LLMs along with the original prompt/query and generate responses which are context-aware and personalized. A few notable fundraising rounds have been announced in the last 3 months — Reka AI (US$58M raise), Contextual AI (US$20M), Langchain (US$10M), and LlamaIndex (US$8.5M).

What is commonly seen recently is the swift embrace of embedding models and vector databases (e.g. Pinecone, Weaviate, Marqo) to enable RAG. Over the past six months, we have witnessed a surge of startups harnessing these capabilities to cater to particular use cases. Examples include providing sales teams with a co-pilot, empowering customer success efforts, enhancing recruiting processes, and enabling advanced data visualization techniques. This wave of innovation signifies the growing recognition and adoption of embedding models and vector databases as powerful tools to drive targeted and impactful solutions across various industries.

Opportunity: Continuous improvement of LLM performance through model validation and observability

As more enterprises understand how to develop and steer LLM capabilities towards expected outcomes, it is crucial to gain insights into the model’s behaviour, performance, and potential biases, especially in critical industries like healthcare and finance where an erroneous output can be literally life threatening. Similar to how SRE and cloud observability vendors became mainstream during the cloud adoption wave, the increasing adoption of AI/LLM will create a demand for observability systems to assess and troubleshoot these models.

By monitoring the inputs and outputs of the LLM, enterprises can identify and rectify any biases or ethical concerns that may arise in the generated content. Observability allows enterprises to detect model drifts, trace root cause and explainability, facilitate debugging, and manage performance (e.g. latency) and costs. Vendors in this category include Arize AI, Fiddler AI, and Helicone. Other companies like Robust Intelligence and Deepchecks are evaluating model effectiveness earlier in the development lifecycle (i.e. shift-left) by providing continuous testing/validation features.

Shadow AI and Data Leakage

LLMs and AI tools are being implemented and used inside organizations, with employees feeding these tools information to elicit responses, which is often confidential company data. However, this has now become a serious cause for concern, because per most AI developers’ terms of service, every time a user prompts a response from LLM tools, they should assume that they are granting permission for their inputs to be ingested into the training dataset and could therefore, be leaked, in verbatim format as outputs for other requests.

Research has shown that at least 4% of employees have input confidential data into ChatGPT. This means that thousands of confidential documents could potentially be at risk of leakage — which could have serious repercussions from a commercial standpoint if these models were to produce outputs that contain private information. To illustrate, when an executive employs ChatGPT to condense minutes from the company’s latest strategy meeting, such information becomes part of the model’s training and can subsequently serve as a resource for addressing future inquiries about the company’s strategy.

As such, several companies are now adopting a stronger stance on the use of AI tools in the workplace — as evidenced by institutions such as Citibank, J.P Morgan and Samsung blocking the use of ChatGPT, while others such as Accenture urge “responsible” use and advise employees to check with managers if and when they plan to use it.

Opportunity: Data access and privacy tools

Recognizing the data leakage issues, LLM and AI tool developers such as OpenAI and Google have come out with some solutions that give users additional controls to opt out/remove their data for model training and output purposes (disclaimer — in some instances):

· Personal data removal — OpenAI has now introduced a Personal Data Removal Request form — mainly for Europe and Japan — for people to request for their information to be removed from OpenAI’s systems. However, the form only points to information being removed from answers ChatGPT provides to users, rather than from its training data. It also disclaims that OpenAI may not, in all cases, delete the data; stating that the company will balance “privacy and free expression” when making decisions about deletion requests.

· Turning off Chat history — Open AI released a setting to allow users to toggle off Chat History and training from a button under their settings which restricts the use of input data for model training and improvement. However, the data is still transmitted and stored by ChatGPT and Bard for “abuse monitoring purposes for up to 30 days” which means that sensitive data may still be vulnerable to leaks should LLM and AI tool developers choose to exploit the data entrusted to them or fall prey to 3rd party malicious actors.

Additionally, opting out as a solution fails to address the lesser studied phenomenon known as “long term memory” of machines, a result of a training method called Long Short-Term Memory (LSTM) neural networks. LSTMs are designed to capture sequential dependencies in data, making them especially adept at retaining and recalling information over extended periods. This means that when an input is introduced to the AI model, there’s a possibility that the model will retain and be able to recall that information in the future, even if the data wasn’t directly used during its training process.

Besides these solutions, some companies have also turned to hosting their own AI and LLMs onsite. For example, in certain cases, specific companies, including those in the financial, military, and government sectors, have no option but to internally host their LLMs. While this approach prevents data leakage by retaining control within the organization, it’s worth noting that internal hosting can be exceptionally costly — with the hardware alone costing north of US$30,000 per unit, and this is on top of the team a company must hire to operate and maintain the model — making it financially unviable for many companies.

Acknowledging the absence of an ideal solution, we feel there is an opportunity for companies to come in and build data privacy/security tools that enable companies to use AI and LLMs more securely. Some of the key solutions our team has been looking at include:

a) Data Masking/Hashing — When users input a prompt that contains private information, the software identifies and masks the sensitive data before sending it to the LLM. The sensitive data sits locally and will be added to the API response for the user consumption, allowing users to extract value from the LLMs in a secure manner. Companies building in this space include PrivateAI, Cado Security’s open source library, and MaskedAI.

b) Browser Isolation — Create secure and isolated browsing environment with configurable policies to regulate the movement of proprietary data while using chatbots or virtual assistants like ChatGPT. An example of such a tool is Surf.

c) Data Loss Prevention — These tools recognize AI and LLM usage in real-time and scans input texts for any sensitive information. When sensitive data is detected, it sends the user an alert and halts/anonymizes the data from being transmitted to the LLM. With LLMShield and CalypsoAI the sysadmins are also alerted about which employee had input sensitive data.

Malicious use cases: Leveraging LLMs to power autonomous threats at scale

Amidst the revolutionary impact of LLMs, it is imperative to acknowledge their vulnerability to abuse. Adversarial users can exploit prompt injection techniques to evade filters and manipulate LLMs into producing deepfake images, videos, and malicious code, even when foundational model companies have embedded prompt firewalls/filters in place to restrict malicious usage. The most common attack vectors using LLMs to date include:

• Spear-phishing — With access to LLMs, bad actors can craft spear-phishing attempts that are highly tailored and personalized, making them harder to detect and resist. Just as how retrieval augmented generation can hyper-personalize content generated for specific audience, the same can be used for spear-phishing campaigns.
• Deepfakes — The rise of multimodal LLMs have brought significant advancements to the field of generative adversarial networks (GANs), allowing malicious actors to generate more realistic and diverse content across different modalities (text, image, video, audio). This enables them to create convincing synthetic profiles that can bypass traditional identity verification (e.g. KYC) and user authentication solutions.
• Polymorphic malware — The demonstration of AutoGPT showcases the capacity of AI agents to reason and execute complex, multi-step commands. This breakthrough not only opens new avenues for innovation but also presents a potential challenge, as adversarial users can leverage these capabilities to design agents that possess the ability to strategize, execute tasks, and adapt dynamically to changing environments.

With the right prompts and context fed to the agent, they are capable of adapting their actions based on the environment in real time (e.g. evading detection systems, sandbox environments, or honeypots) and emulate adversarial strategies used by other hacking groups. Be42Labs has a diagram which illustrates the end-to-end process from reconnaissance to execution. Security vendors including CyberArk and HYAS have demonstrated the use of ChatGPT to create polymorphic malware (i.e. malware that monitors environments and mutates its code/signature to avoid detection).

Opportunity: LLM security/fraud detection tools

Generation and detection techniques are an ongoing cat-and-mouse game. As generation techniques evolve to elude current verification and detection solutions, enterprises need to level up their detection capabilities to match the evolving threat landscape. Some of the approaches we have come across include:

a) Detecting AI-generated text from human-generated text: Several approaches have emerged, including OpenAI’s AI Classifier, DetectGPT, and watermarking.

b) Liveness detection: Technique used to determine whether a biometric trait being presented, such as a facial image or voice recording, is from a live and genuine source or a replica/deepfake. While liveness detection is not new in the industry and has been adopted by numerous fraud detection/authentication players (e.g. Pindrop), we expect to see innovation in approaches that detect anomalies at the metadata layer of images/videos, including pixel distributions, color and texture patterns.

c) Jailbreak prevention: Rather than detecting LLM-generated content/deepfakes, guardrails are placed at the prompting process where users are refrained from manipulating the LLM to generate malicious content. Such examples include Nvidia’s NeMo Guardrails and CalypsoAI Moderator.

As new detection techniques emerge, it is important not to lose sight of existing security technologies including access control, device and network fingerprinting, and behavioral analysis. By capturing and correlating user/device/location activity at multiple points throughout the journey, enterprises can add multiple lines of defenses and dynamically profile risk along the way.

Lack of attribution to content contributors/copyright owners

Base LLMs are pre-trained based on data from the web including Common Crawl, Stack Overflow, and others which apply crawling mechanisms to index billions of data sources in a vast repository of publicly available information. While these data sets have been around for some time, they were originally established for use by academic and non-profit organizations.

In the past, these datasets were considered “fair use” under the Fair Use Doctrine, which implied that their utilization would not infringe upon copyright or intellectual property rights laws. However, as LLM developers now monetize their API services, platforms like Stack Overflow and Reddit have requested compensation in exchange for accessing their data. These platforms have invested significant time and effort in building extensive content repositories while cultivating a loyal user and contributor base, and perceive LLMs as a potential threat that could divert their audience away from their own platforms.

Numerous litigation cases have the potential to act as landmark precedent cases in this respect, including:

• The class-action lawsuit against Microsoft, OpenAI and GitHub for violating their terms of service and copyright law by allowing GitHub Copilot (AI -powered code-generating tool trained on publicly available codes), to reproduce licensed code snippets without providing due attribution.
• Legal action by a trio of artists against Stability AI, Deviant Art and Midjourney accusing them of infringing the rights of artists by scraping images, to train their models, without artist consent.
• Getty Images filing charges against Stability AI for copyright infringements both in the US and UK, in which the latter “chose to ignore viable licensing options and long‑standing legal protections in pursuit of their stand‑alone commercial interests.”

We see two sides to this argument:

From the perspective of model developers, precedent cases like Guild v. Google and Google v. Oracle support the argument that training models on open, public datasets falls under “fair use.” These cases established that transformative use of copyrighted materials, such as Google’s scanning of books or using portions of code to create a new operating system may constitute as “fair use”. In the context of training language models, they could argue that works generated from these datasets are sufficiently transformed to differ significantly from the original sources.

On the other hand, data/content contributors raise valid concerns about potential violations of terms of service by AI developers. For example, platforms like Stack Overflow have “Creative Commons” licenses that require proper attribution when using data from their site. Similarly, Common Crawl’s terms of service specify that users of “Crawled Content” may be subject to separate terms of use from the content owners. Additionally, a key issue here is the risk of redundancy and potential diversion of traffic from these platforms. Language models like GPT-3.5 and ChatGPT can effectively extract information from these datasets, potentially reducing the need for users to visit the original platforms.

Given the possible consequences, one might question whether the use of these datasets can still be considered “fair use” in such cases.

As language models continuously update themselves by scraping data from the web, it becomes increasingly important to address how contributors to content platforms like Github and Stack Overflow are recognized and properly attributed or compensated for their contributions.

Opportunity: Contributor attribution/monetization tools

Currently, the mechanisms for website owners to prevent public facing data from being crawled are either not sophisticated, or they are robust enough but require a great deal of technical expertise and manpower, such as:

• Installing a Robot.txt file — which is a set of instructions for web crawlers that ‘instructs’ them not to crawl your web page(s). However, it is more of a friendly agreement between reputable web crawlers (e.g. Google and Common Crawl) and website owners, and by no means ensures that less reputable web crawlers cannot continue to crawl your site. Additionally, it does not remove any data that has already been crawled from data sets and is far from a measure for future prevention.
• Block web crawling with ‘No index’ — which is a rule set that is used to prevent indexing content by search engines that support the noindex rule, such as Google. However, similar to Robot.txt, it is more of a suggestion for search engine crawlers rather than a hard and fast rule. Moreover, installing no index will prevent your site from being found on search engines — which is a big deterrent for most digital companies.
• Anti-abuse prevention systems — Apply machine/deep learning to analyze user activity and detect anti-scraping bots — LinkedIn has published an article on how they prevent bots from scraping member profile data. This approach typically requires an in-house machine learning team to develop models that can detect behavioural anomalies.

We expect to see new and innovative solutions to these issues through early examples:

a) Shutterstock contributors Fund — As part of this initiative, Shutterstock will place a portion of revenue from the AI content generation tool, as well as external deals for training data with companies such as LG etc., into a creator fund. In doing this, Shutterstock will pay and accrue royalties when an artists’ IP is used in the training of AI-generative models or used for the licensing of generative assets created using Shutterstock’s software.

b) Open source funding startups such as Thanks.dev, which analyses open source projects through GitHub and distributes donations based on contribution activity. Algora.io follows a similar mechanism enabling projects/companies to reward their open source community via bounties.

The need for government regulation

Additionally, besides the need for tools/platforms to support contributor monetization/attribution, we find it important to call out that several of these concerns stem from the lack of standardized regulation in the industry. That said, recent coverage of legal concerns has accelerated initiatives in this domain, and we are currently observing a broad spectrum of approaches being undertaken by different countries — with each adopting a different level of legal intervention.

Globally, the EU is leading AI regulation, having formalized what appears to be the West’s first AI law. Their risk-based regulatory approach classifies AI applications into 4 levels of risk — unacceptable risk (banned by default), high risk, limited risk, and minimal risk. Under this framework, developers of foundation models such as LLMs and generative AI, will be required to apply safety checks, data governance measures and risk mitigations before making their models public. Moreover, AI developers will need to ensure that the training data used to build their models adhere to copyright law — which should help to alleviate several top-of-mind legal concerns.

Similarly, in the East, China is leading the charge with its new, draft AI laws, which aims to balance the growth and security of AI technologies. Specifically, these regulations (which extend to AI applications across sectors) address three key aspects: 1) Generative AI providers must comply with core socialist values, 2) they should refrain from generating illegal or harmful content, and 3) they must register their algorithms with the government. Additionally, developers have to ensure each data source used for AI learning must be accurate, unbiased, and not infringe on IP rights of others. Despite mixed reactions from the public, the new laws signify China’s proactive approach to AI regulation, aiming to foster a controlled and secure AI ecosystem.

What does all of this mean?

Similar to the impact of the internet, cloud, and mobile technologies, the introduction of these advancements has not only sparked remarkable innovation but also presented its own set of risks and challenges along the way. Nonetheless, these challenges have also created fertile ground for 2nd order opportunities (e.g. the rise of cloud adoption has necessitate the need to adopt cloud security).

The figure below illustrates our perspective of the patterns where technological shifts have subsequently paved the way for a proliferation of second-order innovations, many of which have achieved remarkable success, reaching billion-dollar valuations.

We are living in an exciting decade where we are witnessing an unprecedented pace of innovation in artificial intelligence. There is still a lot to be built to realize the vision of placing AI in the hands of everyone in a reliable and responsible manner. That said, we believe this inflection point and the challenges that come along with it — creates whitespaces for new companies to gain and solidify their foothold.

If you are currently building in this space, we would be excited to meet you and find out more.

Disclaimer:

Please note that the views, opinions and analysis set out and expressed in this article are the views and opinions of January Capital only and are not purported to be, or to be construed as any type of legal, accounting, commercial, financial or taxation advice. Please consult with your independent professional advisors to be professionally advised in order to make informed decisions in respect of any relevant subject matter which may have been covered in this article. Under no circumstances is this article or the views expressed (and/or attachments) herein to be used or considered as an offer to sell, or a solicitation of any offer to buy, any security or any interest in any investment vehicle managed by January Capital or any of the relevant portfolio companies invested in by January Capital. This article should not be taken as providing recommendations or opinions regarding the appropriateness or suitability of any investment, investment strategy or sectors for investments. All investments entail risks. Capital and Returns are not guaranteed. While the January Capital believes that this material is correct, no warranty, representation or guarantee of accuracy, reliability or completeness is given or implied and, except for liability under statute which cannot be excluded, no liability for errors or omissions is accepted by January Capital. Further, none of January Capital, nor each of its respective officers, employees, agents and advisers accept, to the extent permitted by law, responsibility for any loss, claim, damages, costs or expenses arising out of, or in connection with, the information contained in this article.

We are excited to announce our investment in Marqo as part of the US$5.2 million Seed round together with Blackbird Ventures and Creator Fund. Based in Melbourne, Australia, Marqo is an open-source multimodal vector search engine that equips enterprises with neural, multi-modal search functionality in their applications. We are thrilled to be partnering with Tom Hamer and Jesse Clark in an opportune time where foundational models are revolutionizing information retrieval and search.

Neural search at its inflection point

In the last 24 months, the world has witnessed an unprecedented pace of innovation and adoption in foundational models and their downstream applications driven primarily by advancements in AI. Within this shifting landscape search is arguably seeing more activity than it has for the last decade, with Microsoft Bing taking on Google, and Google embedding generative AI capabilities in Search at their latest I/O event.

Search and information retrieval has become an integral part of our daily lives and the associated complexities are often taken for granted. Most of us have utilized search multiple times a day, across web browsers, social media apps, e-commerce sites, digital workspaces (e.g. emails), knowledge bases, and many others. Search systems are one of the most prevalent “pick and shovel” technologies embedded in software applications, enabling use cases such as discovery (e.g. e-commerce search), classification (e.g. anomaly detection), natural language processing (e.g. customer support chatbots), and recommendation systems to name a few.

Whilst more sophisticated players such as the likes of tech giants Google and Microsoft have spent millions of dollars and significant resource refining their search functionality, today the majority of applications power their search features with keyword-based search. The downside of keyword matching is that it relies solely on exact matches of search terms without considering the semantic context of the query and documents. If a user searches for “shirt with orange prints”, the system will return exact matches of those keywords in titles and product descriptions, potentially retrieving inadequate or irrelevant results. For example, does the user mean shirt with oranges printed on it, or an orange shirt? Or both? A user searching for “B2B commerce startups” would return results that are explicitly mentioned or tagged as “B2B” and “commerce”. Human language is difficult for computer programs to understand because it is inherently complex, ambiguous, and nuanced. This is where neural search comes in.

Neural search leverages the power of deep learning to understand the semantic meaning of queries and documents. Deep learning models transform documents which contain text, image, or other unstructured data into numerical vector representations, allowing machines to understand the semantic relationships and context between words. Using the previous example, “B2B commerce startups” might be associated with terms like “supply chain”, “marketplaces”, “merchants”, “take rate”, or “contribution margin”.

It is worth noting that word embeddings like Word2Vec and fastText had already started powering natural language processing 5 to 10 years ago. While they are effective to some extent, they were limited in their ability to capture the full meaning and context of natural language. On the other hand, transformer-based models trained on vast amounts of data are able to learn intricate, highly nuanced vector representations of human language. This enables them to create embeddings that are context-aware and capture the meaning of the text at a very granular level of detail.

The proliferation of generative AI applications or LLM-powered “co-pilots” have also created a surge in demand for vector search and database solutions. More solution providers and enterprises realize that crafting more relevant, specific prompts probably isn’t sufficient to generate high-quality responses. Rather than solely relying on prompt engineering, enterprises can also retrieve relevant information from a specific knowledge base, which would make the generated output more informed and contextually appropriate. This is known as retrieval augmented generation (“RAG”), which is emerging as the standard recipe for most generative AI applications.

Marqo’s end-to-end, multimodal approach

Marqo is more than a vector database — they provide end-to-end functionality required to run a production-grade, multimodal vector search engine. With Marqo, developers can build the entire encoding and search pipeline, which includes features like document chunking, encoding data into embeddings, storage, dense retrieval, ranking, and integrations. Developers can bring their own embedding models (e.g. Hugging Face, OpenAI) to transform data into vector representations. Marqo has lexical/keyword search functionality as well, which offers flexibility to developers who want to adopt the hybrid approach and balance the tradeoffs between vector and keyword search (e.g. storage costs).

Furthermore, we believe that the future of search will be multimodal, i.e., a user performs search with text, image, audio, or a combination of these inputs to produce more relevant results. We have already seen multimodal prompting (e.g. MidJourney) for generative AI. This is because describing an item, or an idea in your head from different perspectives helps to generate a more relevant outputs (e.g. describing a red dress via text + showing a similar piece of red dress via image). Marqo provides the infrastructure for multimodal search in an efficient manner. Rather than concatenating text and image embeddings, Marqo takes the approach of combining text and image fields together as one single vector embedding, saving on storage cost.

The use cases for Marqo as a platform are extensive — from e-commerce search to content moderation, fine tuning models for domain-specific tasks (e.g. customer support, sales), fraud detection, and more. To accelerate time to value, Marqo has built verticalized, use case specific features as an extension of their horizontal platform, starting with e-commerce search. For instance, e-commerce enterprises can encode user behavior (e.g. click through rates) to vector representations and assign weightings depending on the type of behavior (e.g. assign higher weightage when the user rates an item vs assign less weightage when a user clicks) — this takes user preferences and past user behaviors into account when recalling and ranking search results for that user. Ultimately this provides a better user experience and in turn greater customer conversion for merchants.

The Team

On top of the platform, Marqo has a solid founding team with strong AI and data background. Jesse Clark has accumulated extensive experience in the field of computational imaging, where he was leading the machine learning team for Robotics AI at Amazon. Prior to this, he was a Principal ML Scientist at Stitch Fix, a publicly traded online personal styling company. Tom Hamer previously worked at AWS as a software engineer where he worked on the Amazon Relational Database Services (RDS) product. We first met the team last year and as we spent more time with them, their focus and approach towards building the platform quickly captivated us.

Progress and future ahead

As an open-source product from day 1, Marqo has garnered much organic attention (>3k GitHub stars) from the developer community and secured several paying customers. In line with this funding announcement, Marqo has also announced general availability of Marqo Cloud, the hosted version of the product.

We are operating in exciting times where large language models are transforming information retrieval and generation, and Marqo is well-positioned to be the infrastructure layer to power the next generation of search and retrieval. We are excited to support the Marqo team in this round and be able to partner with Blackbird Ventures and Creator Fund. If you are a founder building in the LLM infrastructure space, we would love to hear from you.

For more information about January Capital, please visit our website and sign up for our regular newsletter.

For more information about Marqo and how its platform can help embed neural, multimodal search capabilities in your applications, please visit their website and GitHub.

DISCLAIMER

January Capital is a licensed fund manager in both Singapore (through Jan Cap Pte. Ltd.) and Australia (through January Capital Pty Ltd) (together, the “Managers”) licensed under the purview of the Monetary Authority of Singapore and the Australian Securities and Investments Commission respectively.

The information contained in this article is being provided to you for informational purposes only and does not constitute in any circumstance an offer to sell, or a solicitation of an offer to buy any interest in the funds managed by the Managers or any other companies mentioned herein. Additionally, the information in this article and any and all forward-looking statements are based upon assumptions that may not prove to be correct or accurate and the actual financial results/performance and opportunities (amongst others) may differ materially from these statements. As such, nothing contained in this article should be construed as or relied upon as financial, legal or taxation advice, or an encouragement to invest in our funds or any companies mentioned in this article. Past performance is not indicative of future results.

We are excited to partner with Matteo Pelati and Vivek Gudapuri in their journey by investing in Dozer’s US$3.0 million Seed round along with Sequoia Surge and Gradient Ventures. Based in Singapore, Dozer is a data infrastructure platform that enables developers to create blazingly fast data APIs in a plug and play manner. Along with this announcement, Dozer has also emerged from stealth and is now on open beta accessible for developers.

The Emergence of Data Products: Bringing Data to the Frontlines

It has been widely regarded that “data is the new oil” in the 21st century. The “refining” process of data is analogous to that of oil — the need for extraction (i.e. data pipelines), storage (i.e. data warehouse/lake), and processing (i.e. data prep/transformation). To unlock the full potential of this commodity, the last mile is crucial. Similar to oil being refined into products such as gasoline and diesel fit for end user consumption, data sitting in warehouses/lakes can be transformed into data products for downstream use cases (e.g., reporting, analytics, machine learning). A data product is defined as a modular, ready-to-use set of data that users across different teams can leverage to address specific downstream use cases.

In recent times, big data has seen a seismic shift from analytical to operational use cases. The Cambrian explosion of data generated has led to organizations strategizing ways to harness these large troves of data to improve business outcomes. Rather than keeping data in “cold storage”, data and insights are increasingly pushed to frontline systems to facilitate better decision making internally. Functional teams are now able to work with the data in the systems they “live and breathe” in, (e.g., sales teams having 360-degree customer views in Salesforce to enhance account management processes). Vendors that provide this underlying infrastructure known as Reverse ETL and embedded analytics have taken off in the last couple of years.

What if data and insights are not solely exposed to internal stakeholders, but also to end customers as part of the customer/user experience? Consider use cases like displaying real-time price analytics in a stock trading app, visualizing transport location feeds for a fleet management software, or reflecting real-time statuses of manufacturing equipment to minimize manufacturing downtime. These are tangible examples of how data can be embedded in software applications in real time to improve user experience. The response time for these applications often matter, which points to the importance of having real-time capabilities in the underlying infrastructure (e.g. streaming and caching). Data warehouses/lakes designed with OLAP architectures are built for complex and analytical SQL queries and not meant to serve these real time use cases.

Dozer abstracts the complexity required for developers to build and deploy low-latency, highly scalable data APIs

To achieve the above use cases, product and data engineers collaborate to build a “last-mile” data pipeline that allows data to be extracted from a storage system (e.g. database or warehouse), transformed on the fly, and configure an API that can interact with an app. However, building the infrastructure to create and deploy real-time, scalable data APIs is by no means an easy feat. There is a great deal of complexity and data tooling expertise involved, and even then, it is time consuming to configure and manage 4–5 different tools to stitch the entire workflow together (e.g. Kafka, Airflow, Spark, Redis, etc).

Dozer addresses this exact problem and builds on top of storage systems like data warehouses and databases that are not equipped to handle low latency querying and lookups. With Dozer, developers can build the last-mile pipeline — from connecting to the data source to transforming the data on the fly, configuring the APIs and the caching layer, all while providing scale-out and real-time functionality. All of this is done with a single binary which optimizes developer experience, and with a minimal learning curve for product engineers. Dozer supports client libraries in React, Javascript, and Python today.

Dozer is open source and built with Rust — a low-level programming language that is well known for its blazingly-fast performance and memory safety. As more customers demand real-time and edge use cases in future (e.g. AI/ML, real-time cybersecurity, IoT, data streaming, multi-user collaboration), performant developer languages/frameworks such as Rust and WebAssembly will continue to grow in popularity.

The complementary founding team that brings data and product expertise together

Over the last 9 months through numerous sessions with Matteo and Vivek deep diving into the problem they are solving, our level of conviction grew by the day that this is a team with unique insights and complements each other in domain expertise. For one, both co-founders have designed real-time, highly scalable applications and infrastructure in their past stints and have experienced similar complexities first-hand.

With over 20 years of software and data engineering experience, Matteo has been part of data engineering leadership teams and built real-time data platforms at some of the most data-driven yet architecturally complex organizations including Goldman Sachs, DBS, and DataRobot. Vivek comes with deep product and platform engineering expertise, having led technical teams and scaled platforms in fast-growing technology companies like Open Fabric (a January Capital portfolio company), PaySense and Yojee.

Progress and future ahead

Dozer has been heads down building in the last 6 months and today, we are thrilled to announce that they are emerging from stealth and releasing their open beta for developers. Moving forward, they will be focusing on elevating product maturity, hiring, and developer engagement initiatives.

We are excited to be part of Dozer’s early journey and partner alongside Sequoia Surge and Gradient Ventures. If you are a product/data engineer looking to build low latency/streaming data APIs on top of PostgreSQL/data warehouse, please check out their website and GitHub. As a startup that is founded open-source from day one, Dozer is also welcoming contributors proficient in Rust to be part of their contributor community.

For more information about January Capital, please visit our website and sign up for our regular newsletter.

For more information about Dozer and how its platform can hypercharge your databases / warehouses with easily integratable search, analytical, and real-time capabilities, please visit their website and GitHub.

DISCLAIMER

January Capital is a licensed fund manager in both Singapore (through Jan Cap Pte. Ltd.) and Australia (through January Capital Pty Ltd) (together, the “Managers”) licensed under the purview of the Monetary Authority of Singapore and the Australian Securities and Investments Commission respectively.

The information contained in this article is being provided to you for informational purposes only and does not constitute in any circumstance an offer to sell, or a solicitation of an offer to buy any interest in the funds managed by the Managers or any other companies mentioned herein. Additionally, the information in this article and any and all forward-looking statements are based upon assumptions that may not prove to be correct or accurate and the actual financial results/performance and opportunities (amongst others) may differ materially from these statements. As such, nothing contained in this article should be construed as or relied upon as financial, legal or taxation advice, or an encouragement to invest in our funds or any companies mentioned in this article. Past performance is not indicative of future results.

Today, we are announcing our investment in ToolJet as part of the US$4.6 million Pre-Series A financing round led by Nexus Venture Partners with participation from Ratio Ventures and Better Capital. Based in Bangalore, ToolJet is one of the world’s fastest growing open-source startups which offers a comprehensive low-code platform that accelerates the development of internal applications. We are excited to partner with Navaneeth PK in his journey.

As a firm, we invest behind a number of core themes where we can cluster our knowledge. ToolJet represents our 3rd announced investment in the low/no code application development space, together with Locofy and Datature.

Developers spend one to two-thirds of their time building internal applications

Internal applications promote the efficiency of work being done in enterprises/startups today — they are built to (1) handle unify and visualize data from different systems, and (2) address intricate workflows that are inherently unique to an organization. Consider a customer refund process, where a customer service manager needs to look up support tickets (e.g. Zendesk), query customer and payments data (e.g. PostgreSQL/NoSQL database), then process a refund through a payments system (e.g. Stripe) while sending an email/text notification to an end customer (e.g. Twilio/Sendgrid). Rather than interfacing with 4–5 different systems (or more commonly, teams with access to those systems) throughout this process, an internal application can connect all of these systems and serve as a single workspace to visualize data and trigger workflows. The use cases are extensive and have applicability across all functions within an organization, with common ones such as admin panels, reporting dashboards, custom CRMs, and process/approval management workflows.

However, building such internal tools has historically consumed a significant chunk of developers’ time. While the business case to build these highly customized applications for internal adoption is clear (and only becomes stronger as a company scales), there areoften limited development resources dedicated to such initiatives. Instead, developers who were hired to build external, customer-facing product features tend to be asked to invest a percentage of their time into these internal initiatives, in addition to their core role. This often leads to internal conflict as well as a risk of development delays as these internal tools are de-prioritized and fail to be rolled out as planned. Despite this, developers continue to spend (on average) 1/3 of their time building and maintaining internal applications, with those in enterprises larger than 1,000 FTEs spending up to 60% of their time doing so (according to a study done by Retool in 2022).

The market has recognized this sizeable “hair on fire” (US$26 billion) opportunity in the last 3–4 years, with large technology firms developing their own in-house low code application development platforms such as Microsoft Power Apps, AWS Honeycode, Salesforce Lightning App Builder, as well as startups such as Retool.

ToolJet: A developer-first low code platform to accelerate internal app development

Many of these existing tools, however, tend to focus more on the “no code” for non-technical team members — tools that are rigid and are designed to be deployed quickly but with very limited customization. As mentioned in our low/no code landscape primer, as a firm we stand behind low code platforms that put developers at their core — this means adhering to software engineering principles and being interoperable (i.e. no vendor lock-in).

Over the past 12 months, we have been fortunate to get to know Navaneeth and the team at ToolJet. ToolJet leverages the languages and frameworks developers are familiar with and connects with a multitude of backend systems, giving developers a high degree of customizability required to build high-fidelity internal tools at breakneck speed. As we spend more time with the team, we have been impressed by the vision and mission they are focused on.

Today, ToolJet offers “out of the box” UI components and 100+ in-built connectors with databases, SaaS, and cloud storages. Developers can simply connect with their backends in a matter of a few clicks and query/transform the data using Javascript or Python code. The platform is also built to support extensibility, where developers can import their own custom components and connectors if there isn’t one suitable for their use case. For greenfield use cases where developers are building an internal app from scratch, ToolJet provides its in-house database built on top of PostgreSQL. ToolJet also offers “multiplayer mode” that enables real-time collaboration across teams to edit apps concurrently.

For enterprises looking to scale usage across multiple teams and the need to comply with security and compliance standards, ToolJet has a self-hosted deployment option and enterprise-grade features including SSO, RBAC, and audit logging.

ToolJet has amassed developer love and enables developers to contribute easily

At January Capital, we are bullish that open core business models in certain software categories have the potential to capture more value as compared to proprietary software in the long run. It is akin to product-led growth principles (e.g., self-serve, providing value upfront before monetization, iterative product/pricing feedback loops, community driven approach). As feature parity across competitors set in, go-to-market execution becomes more crucial as a value driver of the business. ToolJet is built to be open core from day one as it offers both open source and commercial editions targeting different buyer/user profiles.

Early in a company’s lifecycle, the number of GitHub stars can be a proxy metric to determine the degree of success for any open-source project in terms of adoption and community development. On this metric, ToolJet has grown exceptionally quickly since its initial release since Q2 2021, having gone from 0 to more than 17,000 GitHub stars in less than 18 months (for context, only 2,800+ out of 5.8 million projects in GitHub have amassed more than 10k stars). Having managed to achieve developer love and virality in a world where open-source companies are now ubiquitous, we believe this speaks to the exceptionality and velocity of execution on product and community development by ToolJet.

However, GitHub stars are merely the tip of the iceberg in evaluating open-source projects — building in the open means there is transparency to aspects like code commit frequencies, issues raised and time to resolution, and the level of contributor activity. The other key advantage of being open source besides user adoption, is the opportunity to accelerate product feedback loops and roadmap based on contributions from the developer community. ToolJet has taken a deliberate approach by designing its platform to be contributor friendly with its plugin-based architecture. Developers familiar with Javascript can now contribute by building and publishing custom plugins on top of ToolJet using their plugin development kit (PDK), thereby extending the functionalities of the platform (details can be found in their contributing guide).

To date, ToolJet has contributions from 300+ developers, with 33 of them being active contributors from the community (our definition of ‘active’ is loosely defined as developers excluding ToolJet’s employees who have contributed 3 commits or more in the last 12 months). This number is close to twice of ToolJet’s employee base today.

More importantly, they have achieved this in under 24 months and having only raised US$1.5 million to date, demonstrating both velocity and capital efficiency to achieve early signs of what we termed as “product-community fit”.

Progress and future ahead

ToolJet’s growth trajectory has been primarily organic to date — most developers know ToolJet through word of mouth, forums, and SEO, with many of them already using ToolJet in production environments. We are impressed by the feature/engineering velocity Navaneeth and his team has exhibited in the last 20 months since inception, having recently launched ToolJet v2.0 with a myriad of feature releases and upgrades. Going forward, ToolJet is doubling down its efforts to achieve product-market fit (i.e., retention and monetization) through commercialization and proactive community engagement by institutionalizing DevRel best practices.

We’re delighted to support the ToolJet team in this round, and be able to invest alongside Nexus Venture Partners, Ratio Ventures and Better Capital. If you are a developer who is looking to accelerate internal application projects, do check out their documentation page and GitHub for more details. If you are a founder who is building in the developer tooling ecosystem, we would love to hear from you — we continue to actively review opportunities and deploy capital with more than 10 deals completed in the past 12 months alone, with a core focus on software and B2B opportunities.

For more information about January Capital, please visit our website and sign up for our regular newsletter.

For more information about ToolJet and how its platform can accelerate your internal tooling projects, please visit their website and GitHub.

DISCLAIMER

January Capital is a licensed fund manager in both Singapore (through Jan Cap Pte. Ltd.) and Australia (through January Capital Pty Ltd) (together, the “Managers”) licensed under the purview of the Monetary Authority of Singapore and the Australian Securities and Investments Commission respectively.

The information contained in this article is being provided to you for informational purposes only and does not constitute in any circumstance an offer to sell, or a solicitation of an offer to buy any interest in the funds managed by the Managers or any other companies mentioned herein. Additionally, the information in this article and any and all forward-looking statements are based upon assumptions that may not prove to be correct or accurate and the actual financial results/performance and opportunities (amongst others) may differ materially from these statements. As such, nothing contained in this article should be construed as or relied upon as financial, legal or taxation advice, or an encouragement to invest in our funds or any companies mentioned in this article. Past performance is not indicative of future results.

Our Investment in ToolJet was originally published in January Capital on Medium, where people are continuing the conversation by highlighting and responding to this story.

January Capital recently participated in the $12 million Series A financing round of Twid, a fintech company headquartered in Singapore and India. Twid is building a platform that aims to create a network under the rewards ecosystem that benefits merchants, issuers and customers.

We are thrilled to support Twid on their mission to unlock rewards for all ecosystem participants alongside co-investors such as Rakuten Capital, Google, Sequoia India’s Surge, BEENEXT and others.

Market Opportunity & Thesis

#1 — Market Timing: Fragmented ecosystem of reward points

Reward and loyalty programs are a critical mechanism to attract and retain customers. Creating closed-loop ecosystems help reward customers that demonstrate a commitment to a particular brand or merchant. As e-commerce and digital payments have rapidly increased, merchants and issuers are increasingly focused on providing value via this medium to incentivize and reward the right consumer behaviour.

However, it is not that the number of merchants investing in rewards programs has stagnated as the overall market has grown — the number and type of merchants issuing these points has rapidly evolved. Historically, loyalty programs were primarily offered by financial services institutions (via credit/debit cards), department stores (based on dollars spent in-store), airlines (based on miles flown) or hotels (based on nights stayed). As brands have recognized the importance of such programs, the number of players offering value in exchange for loyalty has exploded. Today, anyone from e-commerce giants (i.e., Flipkart or Nykaa), food delivery players (i.e., Swiggy) or even payment wallets (such as Google Pay and PayTM) offer their own respective versions of rewards.

While these reward programs are designed to benefit customers, the proliferation of digital payments (and earned reward points) has led to hyper fragmentation in the ecosystem. While an increasing number of players offer reward programs, their value is only available to be spent in a “locked” ecosystem. Customers are confined to use these points only within the issuers’ closed ecosystem of merchants, each with their own varying terms and conditions, and based on minimum spend requirements.

So while, customers today earn and accumulate more reward points across many programs than ever before, the perceived value of programs has not met expectations often as customers struggle to utilize their balances. These limited opportunities to redeem value for loyalty makes customers explore other merchants and brands. At the same time, the accounting treatment of points means that points issuers actually do want these points utilized — in addition to building loyalty, they can be a costly line item on a company’s balance sheet.

January Capital believes that this fragmentation will only increase over time and it demands that an intermediary platform create a more “open” approach to rewards monetization, not confined to the closed loop ecosystem of a specific issuer. This is a problem statement that we have been exploring for some years, waiting for the right team to solve this issue.

Twid today is solving this — by building an ecosystem where customers can freely utilize their reward points at any of Twid’s partner merchants (where it aggregates them directly or based on partnerships with different point issuers) as a checkout option.

This problem is a global one — however, we believe Twid’s first market of India is the right testing ground, especially given the domestic addressable market, driven by the proliferation of digital payment acceptance and the presence of the Unified Payments Interface (UPI) network.

#2 — Market Construction: Value proposition that benefits all ecosystem participants

In addition to market timing now being particularly interesting, we have also observed that all ecosystem participants today experience different but equally important pain points today.

Customers

The most common complaint by users pertaining to rewards are that they do not provide actual value — either that a respective program does not offer interesting enough redemption value opportunities or that the guidelines on how to effectively use points is unclear.

Through Twid, customers can view all their disparate reward points accumulated in one single view, and then use their points as a form of currency directly at the final step of their payments page. In essence, this helps with better utilization of their existing points across a broad range of use cases, creating tangible value for consumers.

Merchants

In today’s landscape, merchants are almost always open to new ways to increase customer traffic, to ensure efficient acquisition and retention costs. This is particularly true for localized demand generation platforms — our existing portfolio such as Shopback has benefited from being in-market and offering efficient customer discovery pathways.

By partnering with Twid, merchants are able to gain access to millions of consumers who have billions of points from issuers and can expand their reach to potential customers that may otherwise not normally transact with them. Overall, Twid helps merchants gain access to a larger customer base and issuer pool, demonstrating higher (and quantifiable) ROI than other marketing initiatives.

Issuers

Point issuers, be it banks like or marketplaces, roll out loyalty and reward programs to generate customer engagement. They spend considerable time, capital and effort on these initiatives — both in terms of building partnerships with merchants, as well as marketing these programs/subsidizing these rewards to attract customers. However, their customers are only able to spend these loyalty points in the limited merchant network associated with the point issuers (which is constrained by the amount of resource dedicated to growing this network).

By opening these issuers to a wider merchant base where customers can use their points, Twid is able to drive higher engagement with an issuer’s customers. As such, they are able to achieve their two core objectives of:

1. Engagement: Through repeated transactions and higher spends

2. Profitability: Due to better monetization and in some cases, the presence of merchant funded accruals

#3 — Market Opportunity: Strong partnerships and liquidity enhance value proposition

According to reports, the volume of digital payments in India, is poised to grow to US$1.0 trillion by 2026, reflecting a CAGR of 27.0%. This creates a large and fast-growing opportunity for a player like Twid to empower ecosystem participants to differentiate themselves in a crowded space.

Loyalty programs are also expected to continue growing, with a survey by Global Loyalty Programs estimating potential market size at US$142.6 billion. In addition, the report also mentions that most loyalty/rewards issuers intend to prioritize three key aspects:

Infrastructure, engagement and enhancing the mobile experience are all value propositions that Twid offers today.

Ultimately, Twid’s overarching objective is to increase liquidity in the market, allowing existing reward points to be utilized like a non-fiat currency. This value proposition only gets strengthened by increasing the liquidity of its “marketplace”, allowing for better monetization of the rewards accumulated. As Twid partners with more issuers, acquirers and merchants, it will be able to attract more customers to its platform, in turn then also enhancing its value proposition to issuers and merchants.

January Capital believes that despite being early in its journey, Twid has been successful in generating trust across the ecosystem, with a number of leading merchants and issuers partnered with across India. The company today works with renowned partners such as JioMart, NetMeds, Yatra, Gaana, Wakefit, Ixigo, Payback, Intermiles, PayU, Induslnd and many more. By having a first mover advantage in accumulating these partnerships, Twid can leverage the network flywheel effect to build an effective moat in this space.

Product Overview & Vision

Twid’s platform enables consumers to spend the reward points of issuers at any of the onboarded merchants via a contactless payment mode for offline payments or as a checkout option on online merchants using “Pay with Rewards” function.

When a consumer selects the ‘Pay with Rewards Option’, their tokenized mobile number is passed on to Twid, who uses it as a unique identifier, and pulls together different rewards points systems across their network of issuers, reflecting the consumer’s rewards balance. The consumer can then select which reward they would like to use to make payment. Only one reward can be used per transaction.

Twid manages the full payment volume end to end — both reward points and any remaining dollars not covered by rewards.

The Team

Fintech is one of the key pillars of January Capital’s broader investment thesis on the “digitization of commerce” in emerging markets. Given the complexities involved in building an enduring financial services technology company, we believe founding teams need to have deep “founder-market-fit” to ensure they are able to capably execute on these important but complex opportunities.

From when we first met Twid’s founding team, it became obvious to us that we were engaging with a team capable of building a large platform company. Amit, Rishi and Amit together represent a team that (we believe) are very well positioned to scale rapidly from here.

Twid was only founded in Bengaluru in October 2020 by Amit Koshal, Rishi Batra and Amit Sharma. Amit Koshal has over 15 years of working experience in companies such as Myspace, Balaji Telefilms and ACL Mobile, and prior to his stint as founder, also headed the txtWeb business division at Intuit.

Rishi, meanwhile, prior to his entrepreneurial journey, was part of the Digital Investments team at Bennett, Coleman & Company for over 8 years, where he spearheaded investments in online travel, B2B, E-Commerce and other spaces that are very relevant to Twid today.

Amit Sharma, on the other hand, rounds off this team with strong technological expertise, with over 20 years at the likes of Oracle, Hewlett-Packard, Intuit, HCL and others.

Progress so far and journey ahead

Today, Twid has already partnered with more than 50,000 merchants, who bring a combined 40 million registered end users. With merchants and issuer partners such as JioMart, NetMeds, Wakefit, Ixigo, Intermiles, Induslnd Bank, PayU and others already on board, Twid is looking to expand its offerings and increase overall end liquidity and increase customer delight.

While still early in the journey, Twid has made commendable progress to date, and we are excited to see what the future holds, along with Rakuten Ventures, Google, Sequoia Capital, BEENEXT and others.

Our Investment in Twid was originally published in January Capital on Medium, where people are continuing the conversation by highlighting and responding to this story.

January Capital is a venture capital firm that leverages proprietary data to partner with the most promising founders at the earliest stages. We focus on companies in the Asia-Pacific region who are building technologies that will power the digitization of commerce and the future of work.

Market Perspectives is a series where we summarize the past quarter of venture capital activity in Southeast Asia and offer our perspectives on the trends emerging from this data. We publish this content quarterly, leveraging both internally sourced data as well as information from our partnered company, Venture Cap Insights.

Southeast Asia Deal Tracker — Q1 CY2022

Each quarter, we provide a “tear sheet” that includes a number of key metrics for the Southeast Asian technology company ecosystem by stage of growth. We provide this via our monthly newsletter (interested readers can sign up here). The metrics include:

• Number of rounds: The total number of financing rounds verified via regulatory filings in the quarter;
• Deal value: The total dollar value of financing rounds that formally closed during the quarter;
• Average deal size: The average total size of financing rounds that formally closed during the quarter; and
• Average pre-money valuation: The actual share price paid by each investor in a financing round that closed in the quarter.

Overall, Southeast Asia recorded approximately US$1.7 billion of equity venture financing for Q1 CY2022. This was lower compared to Q4 CY2021, marking a decrease of 36% quarter-on-quarter. Digging deeper, this was largely attributed to the slowdown of Series B (-36%) and Series C+ investments (-43%) in the region. While the first quarter of the year tends to be a slower quarter generally, the total aggregate deal value this quarter is a marked difference to the momentum we observed throughout CY2021.

The most common narrative cited by commentators at present for why this type of slowdown in late-stage growth funding may occur is generally:

(i) A convergence of macroeconomic factors (i.e., geopolitical instability, inflation) are producing a negative environment for “growth” oriented companies;

(ii) Market uncertainty and downward forces have created public market volatility which has resulted in a rapid sentiment change towards “defensive” assets. This has led to substantial re-rating of public market technology multiples, driving down returns for investors; and

(iii) The poor performance of initial public offerings (“IPOs”) and growth stocks generally have left investors unclear on how to price assets, goading a general movement away from the technology sector.

While this narrative is logical, we believe there are also nuances at play in our focus geographies:

• In Q4 of CY2021, more than 40 Series B+ companies raised an aggregate of ~US$2.2 billion of capital (refer to our Q4 market perspectives here). This amount of capital raising was a record quarter for growth stage companies in Southeast Asia and one that felt anecdotally (to us as a firm) as though almost all growth-stage companies were undertaking fundraising processes simultaneously in a single quarter.
• Therefore, it is probably unreasonable to assume that we would see another record successive quarter of fundraising. If we zoom out and look at Q1 CY2022’s trend on an annual basis, year-on-year funding for Series B+ companies is +137% higher than Q1 CY2021, indicating that there continues to be an inflow of patient, long-term capital towards growth assets in Asia.
• This aligns with what we have observed from investment firms raising record funds over the past 12 months — reports estimate that the amount of dry powder held in reserve by GPs today for Southeast Asian technology investments has never been higher (on a dollar basis).
• We can also observe from data held in Venture Cap Insights that at least 33 Series C+ companies have raised successive (two or more) rounds within the last 24–36 months, as timing between funding rounds narrowed over CY2021.
• By capitalizing on this availability of capital during the past 24 months, many of these companies presumably have sufficient cash reserves at present and therefore were not actively in market during Q1 CY2022. Further, we expect that most of these companies could “weather” any mid-term “bear” markets by having cash runway for at least 12–18 months.

At the early stage, investor appetite continues to remain consistently strong — seed stage investments saw a +10% quarter-on-quarter increase in terms of number of financings, and a modest ~3% decline in terms of total funding volumes. It should be noted that alternative financing instruments (i.e. SAFE, Convertible Notes) which are commonly used at this stage of a company’s lifecycle are not represented in our data set. Anecdotally as a firm, we do see an increasing number of financing rounds being completed using these “hybrid” instruments.

Given current price volatility and general market uncertainty, it is reasonable to expect that the use of this instruments increased in Q1 CY2022, meaning overall activity at this stage would be even higher. We analyze the reasons for the continued sustained momentum of Seed funding in our section below on Observed Market Trends.

Observed Market Trends

Early stage activity continues to remain strong even as market volatility persists

Over the past few months, a combination of multiple events including US Federal Reserve rate hikes, record levels of inflation and general geopolitical uncertainty have placed significant downward pressure on publicly listed “growth” stocks. High-growth technology multiples have compressed by almost 50%, according to the Bessemer Cloud Index. Listed companies from Southeast Asia have witnessed similar downward pressure — Sea Limited and Grab Holdings Limited have witnessed CY2022 year-to-date declines of between 45% — 50% as of March 31, 2022.

While there has been a modest turnaround for software companies from absolute market lows (i.e., the Bessemer Cloud Index has increased +22% as of March 31, 2022 from the market low on March 14, 2022), most company share prices remain 30–40% lower than their peak in November 2021. While it is unclear today how public market sentiment will develop in the short-term, for most early-stage investors who tend to adopt a 10-year view when making investments, public market valuations are an important data point but not the sole one.

Despite this volatility in broader markets, data shows that early-stage activity in Southeast Asia continues to remain robust. The number of early-stage financing rounds has remained largely constant over the past five quarters. While there has been a +63% increase in total seed financing volume, this is still much less than broader venture ecosystem which saw a +94% increase year-on-year.

The diverging trends in early and late-stage financing rounds

As we noted above, despite quarter-on-quarter declines for late-stage financings in Q1 CY2022, the year-on-year trend is that deal volume and activity is still up +137% year-on-year. Given the material difference of annual growth in early and late-stage financing activity, we were interested to understand how valuations have changed by stage.

The analysis below shows the actual valuation for financing rounds at the mid-to-late stage (i.e., Series A+) and in the early-stage (which includes Pre-Seed, Seed, and Pre-Series A). This analysis is based on the thousands of financing rounds completed since 2015 in the Asia-Pacific region.

From this analysis, we believe it is possible to draw some inferences:

1. Early-stage investing is more “art” than “science” — using market comparable analysis is only one factor when investors invest in early-stage companies. Despite significant growth in Southeast Asia’s technology ecosystem since 2015, pre-money valuations have remained largely consistent.

2. Dilution may be the key factor in determining early-stage valuations — it is often said a founder’s first financing round is the most expensive. Because of this, founder dilution is a very important consideration for early-stage founders and investors. Interestingly, the median dilution across the 56 deals in Q1 CY2022 in Southeast Asia at this stage was ~20%, which is not materially different than the dilution 5–7 years ago (i.e., round sizes have increased in line with valuations).

3. There may be more volatility for late-stage investing depending on market sentiment — the analysis suggests that during “bull” markets, valuations increase far more rapidly for late-stage than early-stage companies, and in “bear” markets the opposite is traditionally true. Interestingly, despite significant market correction in public markets over the past 3–6 months, pre-money valuations actually increased substantially year-on-year.

4. Given the long timelines, there is limited correlation between early-stage valuations and NASDAQ index returns — a study of the gross returns of over 1,500 seed stage start-ups by AngelList showed that there was zero correlation with the NASDAQ index’s returns. This held true even after accounting for a time lag (to adjust for liquidity in public markets vs. private).

B2B investments now account for a larger share of overall volume

Another interesting observation in Southeast Asia over the past two years has been the continued growth in funding for B2B/B2B2C companies across almost all verticals. From accounting for ~48% of total deals in Q1 CY2021, B2B/B2B2C deals now account for 54% of total deals in Q1 CY2022, underpinning increased investor attention to these companies.

Top Sectors Funded in Q1 2022

Digging deeper into the deal data, four key sectors emerged that saw elevated levels of funding in Q1 CY2022:

• Enterprise SaaS — with a particular focus in cybersecurity, HR, and productivity workflow tools.
• Fintech — in particular, embedded fintech focused on MSME financing solutions.
• E-Commerce — a sustained trend of better servicing and unlocking access to the “long tail” of informal and social commerce merchants.
• Healthtech — mainly tackling genetic testing, telehealth, and mental wellness.

Below are two interesting sectors that that stood out to us:

Social commerce grows into an enormous opportunity

Social commerce has emerged in Southeast Asia over the past one to two years as a particularly attractive segment for venture investors. The affinity with these types of businesses has not slowed — in Q1 CY2022, a number of companies in this space were funded. The emerging opportunity appears to be in:

Social reseller platforms

Particularly those platforms that are building category-specific solutions to empower digital sellers.

Other startups that have previously raised in the space include Kitabeli and Dagangan.

Agent reseller model

The reseller model continues to receive investor interest — interestingly, most of the companies funded in this space were in Vietnam (rather than Indonesia) for the first time.

Other startups that have previously raised in the space include Super, Evermos, Chilibeli and Resellee.

Renewed optimism in embedded financial services that better service SMEs

In an underbanked region like Southeast Asia, alternative credit solutions continue to be critical to continue to support the growth of economies. The rise of direct digital lending services and buy-now-pay-later solutions has fuelled the move away from cash payments and increased consumer usage of ecommerce — online payments in the region are poised to exceed US$1 trillion by 2025.

The first generation of SME focused financing solutions were general in nature — companies such as Koinworks and Funding Societies offer point-solutions that allow SMEs to access much-needed credit for the first time. The challenge with these platforms, however, has been reaching SMEs which are notoriously difficult to identify and assess in the form of credit risk.

Starting 1–2 years ago, we have observed the emergence of embedded fintech solutions that take a “distribution-first” model via products that add immediate value to SMEs and then look to monetize over time. The thesis is that these platforms have real, verifiable data to assess credit risk and can control payment flows to ensure Non-Performing Loans (“NPLs”) remain low. This sector has observed significant funding, with new rounds for the likes of Lummo and Desty closing in Q1 CY2022.

For more information about January Capital and to receive this content going forward, please visit our website and sign up for our regular newsletter.
For more information about VentureCap Insights, please visit the website here.

DISCLAIMER

January Capital is a licensed fund manager in both Singapore (through Jan Cap Pte. Ltd.) and Australia (through January Capital Pty Ltd) (together, the “Managers”) licensed under the purview of the Monetary Authority of Singapore and the Australian Securities and Investments Commission respectively.

The information contained in this article is being provided to you for informational purposes only and does not constitute in any circumstance an offer to sell, or a solicitation of an offer to buy any interest in the funds managed by the Managers or any other companies mentioned herein. Additionally, the information in this article and any and all forward-looking statements are based upon assumptions that may not prove to be correct or accurate and the actual financial results/performance and opportunities (amongst others) may differ materially from these statements. As such, nothing contained in this article should be construed as or relied upon as financial, legal or taxation advice, or an encouragement to invest in our funds or any companies mentioned in this article. Past performance is not indicative of future results.

We are pleased to announce our investment in Locofy.ai’s US$3.0 million Pre-Seed round alongside Accel Partners, Golden Gate Ventures, Bold Capital as well as strategic angels and advisors, including the founders of Ola, Holistic, Hasura, Wego, and GajiGesa. Based in Singapore, Locofy.ai is a low-code tool that accelerates front-end software engineering by automatically translating UI designs to code. We are excited to partner with Honey Mittal and Sohaib Muhammed in their journey to build the most compelling Developer Acceleration as a Service tools available to engineers.

A global problem: the “supply/demand” mismatch for engineering talent

Over the past decade, there has been a creeping issue emerging within the technology ecosystem globally that threatens to slow down the rate of innovation: the increasing deficit between (a) supply of computer science graduates, and (b) the demand for technical talent from technology companies. With the COVID-19 pandemic only accelerating the adoption of digital commerce and software, the availability of technical engineering talent has never been tighter. According to Gartner, 64% of technology executives saw the skill shortage as the most significant adoption barrier to emerging technologies in 2021, compared to just 4% in 2020.

While this is not a new problem per se, historical solutions such as talent offshoring are no longer really viable (with burgeoning local technology ecosystems in markets such as India, Sri Lanka and Vietnam). By 2030, there will be a global shortage of more than 85 million tech workers, and employers are being forced to rethink strategies to ensure they can access the technical talent they need.

In prior investment notes, such as startup oi, we have written about the clear emerging opportunity for startups to better service the talent acquisition component of this problem statement. The reality though is that firms need to explore in parallel how to maximize the efficiency of their existing engineering headcount.

Traditional low/no-code app development tends to be more “disruptive” than “additive” for engineering teams

While no-code empowers a new category of application builders, commonly referred to as “citizen developers”, low code platforms on the other hand have a larger focus towards accelerating development cycles of engineering teams by applying visual development with some coding involved, enabling them to be more extensible and customizable than typical no-code application development platforms.

While there has been an explosion of low/no-code platforms focusing on app development over the past few years, the reality is that very few platforms were built with engineering workflows in mind. These platforms tend to be restrictive in the following ways:

A tendency to to be “closed loop” with limited interoperability

Most traditional low/no-code platforms are closed loop environments and do not work with other developer tools and processes (e.g. version control and CI/CD), resulting in vendor lock-in.

Low/no-code tools are a “black box” for source code manipulation

Traditional tools over-rotate on visual development and the platform does its magic without offering developers the access to the underlying source code, which causes issues if there is an issue with the app and developers have to perform debugging.

There are natural limitations on the scalability of existing platforms

Traditional low/no-code platforms are often seen as a great testing environment (i.e., going from “0 to 1”), but given their rigid nature, once an idea is validated, users are often required to start over from scratch with a traditional engineering process.

For these reasons, traditional solutions tend to not really be “fit for purpose” in professional engineering teams and some have considered it to be more disruptive than additive. Our thesis here is that engineering and product teams require platform agnostic “low-code” solutions that complement their existing software development workflows and help drive efficiencies through workflow improvement and automation.

Locofy.ai: Re-engineering the low-code development process

In modern technology organizations, product managers, designers and developers work together closely to scale new technology products within organizations. Over time, (i) iterative product design, and (ii) effective engineering throughput have emerged as core tenets of successful startups, meaning that effective collaboration between design and development teams has become mission critical.

While the emergence of designer-focused tools (i.e., studios such Figma have created significant efficiencies), the current “design-to-code” process is still far from perfect. Designers and engineers often struggle with unclear processes around how to process feedback, and even after there is alignment internally, thousands of lines of code need to be written to bring the front-end of any application to life.

Locofy.ai is focused on building a highly interoperable, low-code platform that helps supercharge the efficiency of modern software teams — a category of software that the team identifies as Developer Acceleration as a Service (DAAAS).

Locofy.ai’s platform greatly reduces the menial work of engineers without adding complexity or requiring new languages to be learnt (the platform automates 50% or more of the current workflow, starting with front-end code automation). Importantly, Locofy.ai adapts to existing tools and workflows, seamlessly allowing teams to convert their “designs to code” far more efficiently (often standing up a fully interactive and responsive live prototype that runs on code, without writing a single line of code).

Locofy.ai works with the tools and frameworks that designers and developers are familiar with, like Figma and React. Over time, it will look to expand to more design tools and front-end programming frameworks.

A deeply talented founding team with deep experience in the region’s ecosystem

As an early-stage firm that invests from the Pre-Seed rounds, we place considerable weightage on the “market fit” of any founder we partner with and invest into. In the case of Locofy.ai, we are quietly confident that there would not be many other founders who have experienced more “problem statement pain” than Honey and Sohaib have in their careers.

Having worked together for more than a decade, Honey and Sohaib have more than 25 years of relevant experience as a team. They have been part of technical senior leadership at some of the region’s fastest-growing technology companies, including Kredivo, Wego and Homage. At these companies, they have “lived and breathed” the issues they are now looking to solve, and have built up an extraordinary network that they are now leveraging to gain feedback and iterate in an accelerated fashion.

As the Locofy.ai team has grown, we have seen this network also allow them to recruit exceptional talent to join the journey. Locofy.ai’s mission is an important one, and this has resonated deeply with early hires.

Progress and future ahead

After spending most of CY2021 in stealth building the first iteration of the platform, Locofy.ai launched its beta product on December 25, 2021. Since then, usage has grown exponentially — in the past two months alone, early users have generated millions of lines of code across hundreds of projects. This growth was achieved without any deliberate strategies and zero dollars spent to date on marketing — we feel this is a testament to the product-obsessed nature of Locofy.ai’s team and the mission-critical nature of the problem they are building.

If you are a designer or developer looking for a solution to foster collaboration and materially accelerate the “design-to-code” process, Locofy.ai is now in open beta for converting Figma to React code, and accepting early access registrations for React Native, Flutter, Vue, HTML-CSS, and Sketch/Adobe.

If you are an early-stage founder building developer tooling with global ambitions, we would love to speak to you — you can reach our team at contact@january.capital.

For more information about January Capital, please visit our website and sign up for our regular newsletter.

For more information about Locofy.ai and how its platform can supercharge your productivity, please visit their website.

DISCLAIMER

January Capital is a licensed fund manager in both Singapore (through Jan Cap Pte. Ltd.) and Australia (through January Capital Pty Ltd) (together, the “Managers”) licensed under the purview of the Monetary Authority of Singapore and the Australian Securities and Investments Commission respectively.

The information contained in this article is being provided to you for informational purposes only and does not constitute in any circumstance an offer to sell, or a solicitation of an offer to buy any interest in the funds managed by the Managers or any other companies mentioned herein. Additionally, the information in this article and any and all forward-looking statements are based upon assumptions that may not prove to be correct or accurate and the actual financial results/performance and opportunities (amongst others) may differ materially from these statements. As such, nothing contained in this article should be construed as or relied upon as financial, legal or taxation advice, or an encouragement to invest in our funds or any companies mentioned in this article. Past performance is not indicative of future results.

Our Investment in Locofy.ai was originally published in January Capital on Medium, where people are continuing the conversation by highlighting and responding to this story.

We are thrilled to announce our investment in Datature’s US$2.7 million Seed financing round alongside Openspace Ventures. Based in Singapore, Datature is a no-code MLOps platform that enables individuals and teams without machine learning/computer vision expertise to build and deploy computer vision applications. We are excited to partner with Keechin Goh and Denzel Lee in their journey to democratize vision AI.

Flourishing no-code AI market as enterprises look to abstract complexity of building and deploying AI applications

Over the last decade, one the greatest trends in the enterprise space has been the emergence of “no-code” — a programming mechanism that uses a visual development interface to enable less technical users to build software and applications. Designed with a myriad of intuitive and templatized building blocks, no-code takes away the complexity typically required in creating software applications and empowers non-technical teams to become “citizen developers”, instead of relying on resource-constrained IT teams. No-code platforms designed to support many different applications and functional areas have emerged, allowing millions of apps to launch with minimal to no lines of code written — e.g. applications (Appsheet), CMS (Webflow), databases (Airtable), app/data integration (Zapier).

As organizations embark beyond digitization (“Software 1.0”) and adopt data/AI practices (“Software 2.0”) to drive better business outcomes, we see a similar trend where some of the complexity surrounding data/AI can be abstracted away using no-code. In a recent Gartner survey, 87% of respondents believe that investment in AI is expected to accelerate in 2022, and that the lack of skills and product complexity are the main barriers to AI adoption.

Spurred by the growing AI talent shortage, the no-code movement has democratized AI into the hands of non-data scientists/AI practitioners and made artificial intelligence accessible to more organizations. Developers can now potentially create valuable AI use cases at a speed rivaling that of AI practitioners — all without the steep learning curve of data science and machine learning.

No-code AI platforms provide organizations with the opportunity to close the gap by reducing the need to hire costly data scientists and machine learning specialists, while accelerating time to value.

Democratizing access to computer vision — a growing US$12 billion market

Set out to mimic human visual intelligence in machines, computer vision is a field within AI that deals with algorithms trained to see and understand patterns as humans would in the real world. With increasing visual content underlying much of our daily lives and economic activity, computer vision has the potential to unlock massive opportunities with compelling economic upsides.

The use cases are diverse in this US$12 billion market, from inventory optimization and automated checkout in retail, process optimization and quality control in manufacturing, to safety and surveillance in smart cities. Bleeding edge computer vision applications can even power new experiences, such as augmented/mixed reality, autonomous driving, and medical imaging.

Most computer vision use cases today lie in automating repetitive activities at scale and enabling employees to focus on higher-value tasks, with the return-on-investment typically measured with efficiency gains. However, in many cases, the journey in building and operationalizing a computer vision model is riddled with challenging cross-collaboration issues and cross-tool/environment maneuvers. Many companies also lack the in-house expertise needed to build the infrastructure and tooling — and find themselves spending more time learning and navigating across multiple tools, rather than focusing on building.

Datature’s end-to-end solution

As we sat down with the Datature team to understand how they are approaching the market and delivering a competitive product global from day one, several key highlights emerged.

Firstly, Datature’s no-code, end-to-end platform allows “citizen data scientists” to be fully immersed in the model training and deployment experience without having to switch between environments. The platform is also versatile and interoperable without vendor lock-in, implying that users can utilise Datature’s ecosystem at any part of their workflow.

Secondly, Datature believes that an accurate computer vision model requires high quality image datasets, and poor data inputs would result in “garbage-in, garbage-out”. To that end, Datature’s AI-powered data labeling tool, Intellibrush, enables users to outline complex objects in a more efficient manner. The platform also has data augmentation capabilities that increases the diversity of images within the training dataset.

Finally, the team is heavily invested in making the platform’s UI/UX sleek and approachable for less technical users.

The Team

Datature’s co-founders first met when they were serving the Singapore Armed Forces as part of the 1st Commando Battalion in 2013. The inspiration to build Datature came from their passion in computer vision and after winning the Rolls Royce Data Challenge in 2019, when they had custom-built a computer vision use case for Rolls Royce. Soon after, they realized the gap between the vast potential of computer vision applications and the challenges in building and operationalizing these applications, and set off to build a platform that could address this problem at scale. As a bootstrapped venture to date, they have managed to build robust features and strong clientele traction — showing the team’s incredible grit, focus and potential for future growth.

Progress so far and journey ahead

Datature has a broad set of existing clientele ranging from construction, aviation, and education sectors, and is looking to supercharge its go-to-market efforts in the region and beyond by (1) enabling a self-serve experience where users can sign up via its website, and (2) targeting mid to large organizations with a clear business case and tangible ROI from deploying computer vision.

We pride ourselves on partnering with founders that are building the early innings for a meaningful and defensible breakthrough in deep technology, and Datature is a prime example of such. With an ambitious team, opportune market timing and opportunity as well as a great product, we look forward to supporting Datature’s ambitious journey to democratize computer vision to the masses and augment business decision makers with AI-driven visual analysis.

The new funding round will further accelerate strategic hires, so please check out their job board to join the Datature family.

For more information about January Capital, please visit our website and sign up for our regular newsletter. If you’re a founder helping to solve a problem within the realm of data and AI, we would love to have a conversation — drop us a message at contact@january.capital.

For more information about Datature and how its platform can support your company’s employee engagement, please visit their website.

DISCLAIMER

January Capital is a licensed fund manager in both Singapore (through Jan Cap Pte. Ltd.) and Australia (through January Capital Pty Ltd) (together, the “Managers”) licensed under the purview of the Monetary Authority of Singapore and the Australian Securities and Investments Commission respectively.

The information contained in this article is being provided to you for informational purposes only and does not constitute in any circumstance an offer to sell, or a solicitation of an offer to buy any interest in the funds managed by the Managers or any other companies mentioned herein. Additionally, the information in this article and any and all forward-looking statements are based upon assumptions that may not prove to be correct or accurate and the actual financial results/performance and opportunities (amongst others) may differ materially from these statements. As such, nothing contained in this article should be construed as or relied upon as financial, legal or taxation advice, or an encouragement to invest in our funds or any companies mentioned in this article. Past performance is not indicative of future results.

Our Investment in Datature was originally published in January Capital on Medium, where people are continuing the conversation by highlighting and responding to this story.