<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by JJ on Medium]]></title>
        <description><![CDATA[Stories by JJ on Medium]]></description>
        <link>https://medium.com/@yongjiajun?source=rss-904928cf7edb------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*AtwiLD0XIS8kxMBgdLdpkg@2x.jpeg</url>
            <title>Stories by JJ on Medium</title>
            <link>https://medium.com/@yongjiajun?source=rss-904928cf7edb------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 06:31:31 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@yongjiajun/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[124th Monthly Technical Session]]></title>
            <link>https://medium.com/henngeblog/124th-monthly-technical-session-8764b30a879f?source=rss-904928cf7edb------2</link>
            <guid isPermaLink="false">https://medium.com/p/8764b30a879f</guid>
            <category><![CDATA[hennge]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[monthly-technical-session]]></category>
            <dc:creator><![CDATA[JJ]]></dc:creator>
            <pubDate>Fri, 21 Mar 2025 02:01:29 GMT</pubDate>
            <atom:updated>2025-03-21T02:01:29.329Z</atom:updated>
            <content:encoded><![CDATA[<p>HENNGE’s 124th Monthly Technical Session (MTS) took place on November 29th, 2024, bringing together members for an engaging exchange of technical knowledge.</p><p>Here’s a glimpse into the insightful presentations that filled the day…</p><h3>Kelvin: Introduction to Local LLMs</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*W4qJxw3rGsJJ0pJzl79T6A.jpeg" /><figcaption>Kelvin presenting <em>Introduction to Local LLMs</em></figcaption></figure><p>Kelvin kicked off the session with an insightful dive into the world of local Large Language Models (LLMs). He first outlined the challenges of working with LLMs, particularly the high costs associated with training and the limitations of their knowledge base. To address the issue of outdated information, Kelvin introduced Retrieval-Augmented Generation, a clever technique that provides context to the LLM during inference, effectively expanding its knowledge in a cost-effective way.</p><p>The presentation then shifted focus to the advantages of running LLMs locally. Kelvin highlighted benefits such as enhanced data control, potential cost savings, the ability to fine-tune models for specific needs, and greater flexibility with licensing. He explained that running an LLM locally requires two key components: an LLM Inference Server and a suitable model.</p><p>Choosing the right model is crucial, and Kelvin provided guidance on key factors to consider. He emphasized the importance of evaluating the number of parameters, which directly affects accuracy and resource requirements. Quantization, a method for reducing model size and resource consumption, was also discussed. Kelvin further touched on “instruct variants,” which are models specifically fine-tuned for chat and dialogue applications, and the importance of understanding licensing options (open-source, research, or proprietary).</p><p>In addition to the core concepts, Kelvin briefly explored alternative approaches to running LLMs locally. He mentioned different inference servers like vLLM, LLaMA++, and MLX, and highlighted valuable resources like Hugging Face, Model Zoo, and Kaggle for finding and experimenting with various models.</p><p>To bring these concepts to life, Kelvin concluded his presentation with a captivating demo using <a href="https://ollama.com/">Ollama</a>, a lightweight and extensible framework designed for building and running LLMs on local machines.</p><h3>Jonas: The Busiest Train Stations in the World</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1TRcZZCxq80Km7b6logyDQ.jpeg" /><figcaption>Jonas presenting <em>The Busiest Train Stations in the World</em></figcaption></figure><p>Jonas’ presentation, aptly titled “<em>The Busiest Train Stations in the World</em>,” offered a compelling exploration of data interpretation and the critical thinking required to avoid common pitfalls. He challenged the audience to question seemingly straightforward data points, using the popular example of “busiest train stations” to illustrate how easily misinterpretations can arise.</p><p>While Shinjuku Station is often cited as the world’s busiest, Jonas highlighted the inconsistencies in defining and measuring “busyness.” Different metrics, such as passenger boardings, transfers, and alightings, are not always uniformly considered, leading to potentially misleading rankings. He further pointed out the complexity of Shinjuku Station itself, which comprises multiple interconnected stations, adding another layer of ambiguity to the definition of “busiest.”</p><p>To further emphasize the importance of context, Jonas examined SaaS usage data (specifically messaging app usage) at HENNGE. He demonstrated how raw data, when devoid of context, can lead to inaccurate conclusions and ultimately misguided decisions.</p><p>The key takeaway from Jonas’ presentation was clear: data is not inherently objective and must be interpreted within its specific context. Data alone is insufficient for informed decision-making. Instead, a comprehensive understanding of the data’s origins, collection purpose, completeness, and potential biases is crucial for drawing accurate and meaningful conclusions.</p><h3>Habibi: Know Your Boundaries</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bD-_v7QJGnb4n8GSLSRu8g.jpeg" /><figcaption>Habibi presenting <em>Know Your Boundaries</em></figcaption></figure><p>Habibi’s presentation delved into the critical role of <a href="https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies_boundaries.html">IAM Permissions Boundaries</a> in securing AWS environments. He began by clarifying the distinctions between IAM Roles, Permissions, and Policies, laying a solid foundation for understanding how access control is managed within AWS. The core concept of Permissions Boundaries was then introduced: they act as a safeguard, limiting the maximum permissions an IAM role can possess, even if overly permissive policies are attached.</p><p>While existing security measures might seem sufficient at first glance, Habibi emphasized the unique value of Permissions Boundaries in preventing privilege escalation attacks. He illustrated this with a common scenario in CI/CD pipelines, where a compromised deployment role could potentially create new roles with unrestricted access, posing a significant security risk.</p><p>Permissions Boundaries effectively mitigate this risk by establishing a ceiling on permissions for any roles created by the deployment role. Even if the deployment role is compromised, it cannot create roles that exceed the restrictions defined in the boundary policy.</p><p>The presentation provided a technical deep dive into the mechanics of Permissions Boundaries, including example policy structures for both the child role boundary (applied to created roles) and the deployment role boundary (applied to the role creating other roles). Habibi demonstrated how to configure the deployment role’s policy to enforce the attachment of the child role boundary, effectively preventing privilege escalation.</p><p>The key takeaways underscored the importance of implementing Permissions Boundaries in CI/CD pipelines that involve IAM role creation. Habibi encouraged the audience to actively explore and leverage the security features offered by AWS, even those whose value might not be immediately obvious.</p><h3>Nicky (GIP): 1-bit LLMs</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PnINZIXOsK5WYJ6Kl4_0VQ.jpeg" /><figcaption>Nicky presenting <em>1-bit LLMs</em></figcaption></figure><p>Nicky’s presentation explored the fascinating world of 1-bit Large Language Models (LLMs), a cutting-edge optimization technique designed to tackle the resource challenges posed by traditional LLMs. LLMs, like the renowned GPT-3, are powerful neural networks trained on massive text datasets to generate human-like text. However, their full-precision weights and sheer size demand significant storage and energy, making them difficult to deploy on resource-constrained devices.</p><p>1-bit LLMs offer a solution by compressing model weights into compact binary representations (1-bit). This clever technique dramatically reduces memory requirements and computational complexity. It achieves this by replacing high-precision values with binary approximations (-1 or +1), while employing a scaling factor to maintain the dynamic range of the model.</p><p>This innovative approach brings several advantages, including enhanced memory efficiency, faster computation speeds, reduced energy consumption, and improved scalability. However, 1-bit LLMs also present challenges, such as potential accuracy loss, complexities in fine-tuning, and hardware constraints.</p><p>Despite these challenges, 1-bit LLMs hold immense potential for a variety of applications. Nicky highlighted their use in gradient compression, creating compact models for efficient inference and enabling real-time AI applications.</p><p>Looking ahead, research in 1-bit LLMs is focused on exploring hybrid quantization techniques, developing improved algorithms, and designing specialized hardware to further enhance their capabilities. Nicky emphasized that 1-bit LLMs represent a promising optimization strategy with the potential to make AI more accessible, efficient, and sustainable.</p><h3>Hans: AI Application in Sports</h3><p>Hans, a self-proclaimed sports enthusiast, delivered an engaging presentation that explored the emerging field of AI-driven sports analytics. He highlighted how the increasing availability of data and rapid advancements in AI are transforming the way teams analyze player performance, predict game outcomes, and make strategic decisions.</p><p>Drawing inspiration from the “Moneyball” approach to baseball, which famously used statistical analysis to identify undervalued players, Hans demonstrated the versatility of machine learning across various sports. Using readily available datasets with over 100 attributes, he showcased how techniques like linear regression and custom neural networks can predict in-game events, such as whether the next play in an NFL game will be a run or a pass, with remarkable accuracy. This approach, he explained, can be extended to player evaluation, predicting game or season winners, and even real-time play prediction.</p><p>The presentation then delved into more sophisticated applications of AI in sports. Hans discussed how LSTM neural networks, a type of deep learning model, can be used to evaluate hockey player value by analyzing historical performance data. He also explored how AI can be trained to identify different play types in various sports, paving the way for more nuanced analysis and strategic planning. Furthermore, he emphasized the potential of AI-powered vision processing to track player and ball movements in real-time, enabling automated detection of events like offsides in soccer or out-of-bounds calls in tennis and badminton.</p><p>Finally, Hans touched on the exciting possibilities of personalized athlete training using AI. By capturing and analyzing detailed player movement data, AI systems can provide tailored feedback and suggest specific areas for improvement. He also illustrated how AI can be used to optimize in-game strategy by calculating win probabilities for different scenarios, such as deciding whether to go for a field goal or attempt a first down in American football.</p><p>Keeping with tradition, the MTS concluded with a beer bash, a chance for attendees to connect and relax in a casual setting.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lACKxjG2w3uu4YFeD85Gow.jpeg" /><figcaption><strong>Kanpai! 🍻</strong></figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8764b30a879f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/henngeblog/124th-monthly-technical-session-8764b30a879f">124th Monthly Technical Session</a> was originally published in <a href="https://medium.com/henngeblog">henngeblog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[96th Monthly Technical Session]]></title>
            <link>https://medium.com/henngeblog/96th-monthly-technical-session-584d637c9ba9?source=rss-904928cf7edb------2</link>
            <guid isPermaLink="false">https://medium.com/p/584d637c9ba9</guid>
            <category><![CDATA[golang]]></category>
            <category><![CDATA[software-engineer]]></category>
            <category><![CDATA[monthly-technical-session]]></category>
            <category><![CDATA[japan]]></category>
            <category><![CDATA[nft]]></category>
            <dc:creator><![CDATA[JJ]]></dc:creator>
            <pubDate>Fri, 14 Oct 2022 02:32:03 GMT</pubDate>
            <atom:updated>2022-10-14T02:32:03.808Z</atom:updated>
            <content:encoded><![CDATA[<p>MTS, short for Monthly Technical Session, is held at HENNGE every month as a platform for everyone at the company to share their ideas and knowledge in a mini-conference-style internal session.</p><p>On 22 July 2022, we held the 96th MTS in a hybrid fashion. Half of the speakers presented on-site at the Shibuya Headquarters, while the rest were interns presenting remotely during their participation in the <a href="https://medium.com/henngeblog/tagged/gip">Global Internship Program (GIP)</a>.</p><h4>Introduction of AI Avatar and Voice Cloning Technology</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-1ToUK7YK7Pve8UHTT6kVQ.jpeg" /><figcaption><strong>Ayaka-san</strong> presenting <strong>Introduction of AI Avatar and Voice Cloning Technology</strong></figcaption></figure><p>First up, Ayaka-san walked us through how the technology of AI avatars and voice cloning can be applied creatively at work.</p><p>Since Business Planning and Analysis Division is in charge of preparing the materials for the quarterly financial reports, they took this opportunity by leveraging the power of such technology with <a href="https://www.synthesia.io/">Synthesia Studio</a> and <a href="https://www.descript.com/">Descript</a> to create the following <a href="https://www.youtube.com/watch?v=qnXwGQqPBFg">Q2 FY22 Earnings Results</a> presentation:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FqnXwGQqPBFg%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DqnXwGQqPBFg&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FqnXwGQqPBFg%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/0f4dd4581683e7827299d7db729e87e6/href">https://medium.com/media/0f4dd4581683e7827299d7db729e87e6/href</a></iframe><blockquote>“We would like investors to understand not only our financial results but also our growth potential. To do so, we need to seek opportunities to realize HENNGE’s value, “Take the First Bite” in our Investor Relation activities such as this one.”</blockquote><p>Later in the Q&amp;A session, Ogura-san mentioned that it was saving both him and Amano-san time and effort not having to block out their busy schedule to run through the script in person.</p><h4>CPRD Talent Mapping Questionnaire (June 2022)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MW8BzgL2bH1rbJLfY2mJqg.jpeg" /><figcaption><strong>Iskandar</strong> presenting <strong>CPRD Talent Mapping Questionnaire</strong></figcaption></figure><p>Next up, we had Iskandar go through the questionnaire results conducted within the Cloud Product Development Division (CPRD/engineering) for talent mapping in the future and also to learn more about the engineers.</p><p>The survey focuses on various aspects of engineer life in HENNGE, such as:</p><ul><li>How engineers rate their skills/expertise (languages/frameworks)</li><li>What skills engineers are interested in picking up</li><li>Project-related preferences</li><li>Interests in activities outside of engineering (OSS, certifications, etc.)</li></ul><h4>Learning with Block-based Programming</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F09UhQoAd8xn9BeZ4-lXZA.jpeg" /><figcaption><strong>Jasmine</strong> presenting <strong>Learning with Block-based Programming</strong></figcaption></figure><p>Moving on, Jasmine talked about learning programming with Block-based Programming based on her prior tutoring experience.</p><blockquote>“Block-based programming (BBP) languages are getting increasingly common these days as an entrypoint to programming for young learners. It allows users to “write” code by dragging and dropping graphical blocks.”</blockquote><p>Out of the many BBP languages out there, we learnt that <a href="https://scratch.mit.edu/">Scratch</a> is one of the more popular ones and you can do a lot with it such as making games and animations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZOPBl9-W02LcVXOh" /><figcaption>Programming concepts through blocks on Scratch: Recursion</figcaption></figure><p>While BBP serves as a good starting point for anyone who’s just getting started with programming, some parts of it do not translate directly into text-based programming and may encourage bad software programming practices if they want to get serious about it.</p><h4>Switching to Declarative UI</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cNDGbXfOo8vOxVFLJ7IsJQ.jpeg" /><figcaption><strong>Charles</strong> presenting <strong>Switching to Declarative UI</strong></figcaption></figure><p>Charles, who specialises in mobile app development then gave a talk about the switch of mobile-based programming languages moving towards Declarative-style UI.</p><p>For context, here’s a quick comparison between Declarative UI (modern-style UI) and Imperative UI (old-style UI):</p><p><strong>Declarative UI</strong></p><ul><li>UI is generated when needed.</li><li>The layout is dynamic and combined with the data population.</li><li>When data changes, UI is completely redrawn instead of updated.</li></ul><p><strong>Imperative UI</strong></p><ul><li>Is a relatively static, persistent UI.</li><li>The layout is generally separated from the data population.</li><li>Changes are performed on the existing UI at runtime.</li></ul><p>While there are upsides to using Declarative UI, such as:</p><ul><li>Faster/Easier to write</li><li>Better cross-platform design</li><li>Another way to force Model/UI separation</li></ul><p>… it’s not perfect, some drawbacks include:</p><ul><li>Requires a full rewrite of UI</li><li>Some widgets are not supported yet</li><li>Some retraining of developers may be required</li></ul><p>Charles then walked us through a couple of case studies on creating sample Android and iOS UI components, the problems we may encounter and the solutions to work around them.</p><h4>Microsoft Power Platform: The Future of App-Building</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AKwipd43KnPVBJA3xg4P8A.png" /><figcaption><strong>Hiroon</strong> presenting <strong>Microsoft Power Platform</strong></figcaption></figure><p>After a short break, Hiroon talked about <a href="https://powerplatform.microsoft.com/en-us/">Microsoft Power Platform</a> and how it is the future of app-building.</p><p>It is a low-code development platform that aims to reduce the need for IT in an organization with a powerful set of applications:</p><ul><li><strong>Power BI</strong>: a business analytical tool that enables users to connect with several data sources, modify, shape and transform data according to the reporting needs</li><li><strong>Power Apps</strong>: an application development tool for building cross-platform applications using <a href="https://docs.microsoft.com/en-us/power-platform/power-fx/overview">MS Power-FX</a>, a low-code/general-purpose programming language.</li><li><strong>Power Automate</strong>: a process Automation tool that allows the automation of manual repetitive tasks.</li><li><strong>Power Virtual Agents</strong>: a tool that allows building intelligent virtual chatbots without having to write a single line of code.</li></ul><p>With some of the use cases presented, Hiroon then elaborated on how it could be the future of app building by going through some of the benefits it could bring when utilised effectively.</p><h4>NFT and the Rise of Generative Art</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lX1EU5267knrqmmR3kFKjQ.jpeg" /><figcaption><strong>Chaw Chit</strong> presenting <strong>NFT and the Rise of Generative Art</strong></figcaption></figure><p>Next up was Chaw Chit presenting NFT and the Rise of Generative Art.</p><p>NFT, short for non-fungible token, is a record on a blockchain with unique metadata and IDs that differentiate them from one another to serve as a proof of ownership and authenticity of digital assets.</p><p>With the rise of NFTs comes the rise of generative arts, where arts are programmed using a computer that intentionally introduces <strong>randomness</strong> as part of its creation process.</p><p>With a little bit of coding, you can start creating your own NFTs by creatively generating arts as simple as mathematical artworks…</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GCqaxA-C0tF8q_KsaLi1ww.jpeg" /><figcaption>Mathematical Artwork</figcaption></figure><p>… or get funky by building your own NFT generative art collection:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oUWN7t883-mECm8u_8qerA.jpeg" /><figcaption>NFT Generative Art Collection</figcaption></figure><h4>Machine Learning with Golang</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tVOCIgXxmK_OHU3UnZBQpw.jpeg" /><figcaption><strong>Camille</strong> presenting <strong>Machine Learning with Golang</strong></figcaption></figure><p>Moving forward, we had Camille give a presentation on machine learning with Golang.</p><p>Given that we have the option to choose a more popular and mature machine learning framework with Python such as <a href="https://www.tensorflow.org/">Tensorflow</a> and <a href="https://pytorch.org/">PyTorch</a>, why should we be considering something else then?</p><p>It ultimately boils down to the following issues with Python: <strong>debugging issues</strong> and <strong>the lack of static typing</strong>.</p><p>With Golang, static typing is possible while its error handling mechanism is more intuitive and simpler compared to its counterparts, which should offer better maintainability when we want to deploy to production at scale.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IDwfx4qfCoOS9vMf91_0rQ.jpeg" /><figcaption>Overview of Go Machine Learning Packages</figcaption></figure><p>Camille then presented her sample code in Golang for building a machine-learning data model with KNN algorithm and using it for image recognition.</p><h4>Bushfire Evacuation System</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iC-2AuVXYq-NZKpgkp8noQ.jpeg" /><figcaption><strong>Olivia</strong> presenting <strong>Bushfire Evacuation System</strong></figcaption></figure><p>Last but not least, we had Olivia present Bushfire Evacuation System for her final year project.</p><p>Australia is prone to bushfires due to its extreme climate in the summer and with the dominance of Eucalyptus trees around the country, their natural oil only further adds to the already perfect recipe for a wildfire.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ij16P3PDyO5G9YkogZ_laA.jpeg" /><figcaption>Bushfires in Victoria, Australia</figcaption></figure><p>The bushfire evacuation system aims to provide details on how people usually move around in regional locations, how hazardous flames spread, and how to improve the safety of the locals.</p><p>Adding onto an existing bushfire evacuation simulator, Olivia and her team produced a user dashboard that enables quick and efficient comparisons of multiple complex evacuation scenarios, aiming to be easily picked up and understood by users, especially volunteers who often may not have knowledge in firefighting.</p><h3>Beer Bash 🍻</h3><p>Following the tradition at HENNGE, we had Beer Bash on site while adhering to the COVID-19 protocols 🍣🍕🍻!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IgbrgfdalLnuoQ8u6ametg.jpeg" /><figcaption>Ogura-san Speaking at Beer Bash after the 96th MTS</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*321gs8bQXf2whBeXcBTD6Q.jpeg" /><figcaption><em>Kanpai/Cheers!</em> 🍻</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=584d637c9ba9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/henngeblog/96th-monthly-technical-session-584d637c9ba9">96th Monthly Technical Session</a> was originally published in <a href="https://medium.com/henngeblog">henngeblog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>