<?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 SHIVANGI SINGH on Medium]]></title>
        <description><![CDATA[Stories by SHIVANGI SINGH on Medium]]></description>
        <link>https://medium.com/@shivangibitsp?source=rss-e0fe6055f682------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*Wokpz1iPEnZH5lM2.</url>
            <title>Stories by SHIVANGI SINGH on Medium</title>
            <link>https://medium.com/@shivangibitsp?source=rss-e0fe6055f682------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 14:16:47 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@shivangibitsp/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[Shopping Emily’s wardrobe: A practical look at Flam’s Flicks]]></title>
            <link>https://medium.com/@shivangibitsp/shopping-emilys-wardrobe-a-practical-look-at-flam-s-flicks-4cf65a51b63b?source=rss-e0fe6055f682------2</link>
            <guid isPermaLink="false">https://medium.com/p/4cf65a51b63b</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[interactive-content]]></category>
            <category><![CDATA[flam]]></category>
            <category><![CDATA[social-media-marketing]]></category>
            <category><![CDATA[digital-marketing]]></category>
            <dc:creator><![CDATA[SHIVANGI SINGH]]></dc:creator>
            <pubDate>Thu, 17 Sep 2026 17:37:43 GMT</pubDate>
            <atom:updated>2026-09-17T17:38:26.219Z</atom:updated>
            <content:encoded><![CDATA[<p>People don’t watch <em>Emily in Paris</em> only for the plot. A big part of the appeal is the clothes: the bright coats, the clashing prints, the bags that seem to change every scene. Every season, fan accounts on Instagram and TikTok go through episodes frame by frame to work out what Emily, Sylvie and Mindy are wearing and where to buy it.</p><p>The show already works like a fashion magazine that moves. The difference is that a magazine prints the brand and price next to the photo. With a streaming show, you pause, screenshot, search, and hope someone has already identified the jacket. Even when they have, it’s often a designer piece most viewers won’t spend that much on, and finding a cheaper version that looks similar is another round of searching.</p><p>Right now, getting from an outfit on screen to comparing similar looks, saving one and buying it takes several apps and a lot of searching. Flam, an AI company that builds interactive video for brands, is working on exactly this kind of problem with a format called Flicks. It raised a $40 million Series B led by QED Investors in September 2026.</p><h3>What Flicks are</h3><p>Flicks are interactive content units that holds several versions of itself. While it plays, a viewer can tap or swipe to change a person, a product or the background, and Flam says the switch takes under 10 milliseconds, so it doesn’t look like the video is reloading. Flam aims it mainly at product pages, where shoppers flip through colours and variants, and at digital ads, where the goal is to get people to interact instead of scrolling past.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8AvTYe6oiAy_4--3-ZCw4Q.png" /><figcaption>Flicks being used for interactive digital ads &amp; product pages</figcaption></figure><h3>How Flicks work</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iU8YnCdUDxAPHJ7Kc9ywaQ.png" /><figcaption>Flicks: Under the hood</figcaption></figure><p><em>Flam hasn’t published a detailed technical breakdown. Based on its product pages and announcements, it works in roughly five steps.</em></p><p><strong>1. Creating the versions.</strong> Flam has an editing model called Forge that replaces objects, characters or backgrounds in a video. It works from reference images and changes only the element you choose, leaving the rest of the scene as it was. Instead of shooting a scene three times with three coats, a brand could shoot once and generate the other two.</p><p><strong>2. Packing them into one file.</strong> All the versions are compressed together into a single file instead of being stored as separate videos. Flam says this makes the file about 60% smaller, and the compression method is patented. Smaller files matter because heavy video slows pages down, and slow ads get skipped.</p><p><strong>3. Delivering it.</strong> Flicks start playing in about 50 milliseconds and can be served programmatically, which means an ad system can decide which version a particular viewer sees. They can be embedded on a website, sent as a link through ads, social media, email or SMS, or opened by scanning a Flam code on a billboard, a TV screen or a shop display.</p><p><strong>4. Switching.</strong> The viewer taps, swipes or hovers, and the video changes to another version without stopping. The two speed numbers measure different things: about 50 milliseconds to start, under 10 milliseconds to switch once it’s playing.</p><p><strong>5. Recording what people pick.</strong> Flam says Flicks track which versions people choose, how often they interact, how many times they switch in a session, and how quickly they first engage. A regular video ad tells a brand how long someone watched. Flicks tell them which option the viewer went for, and that can shape the next set of versions.</p><h3>Applying it to Emily in Paris</h3><p><em>This is my own example. There’s no indication Flam, Netflix or the show’s producers are working on anything like it.</em></p><p>Here’s how a first version could work.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Blkrejdh7h3pZVYfUrpi1Q.png" /><figcaption>MVP for Flicks integrated on a show like ‘Emily in Paris’ on Netflix</figcaption></figure><p><strong>Tap what you’re looking at.</strong> Emily walks down a street in a green coat. Small markers sit on the coat and the bag. Tap the coat and a panel slides up showing what she’s wearing: the label and the price. That alone would save fans a lot of screenshotting.</p><p><strong>See cheaper options in the same scene.</strong> The panel lists three versions: the piece as worn, a similar mid-range option from a partner brand, and a budget pick from a high-street brand. Tap one and Emily is wearing it in the scene, on the same street, in the same light. You see how the cheaper coat actually looks on screen, which is more useful than a product photo on a white background. Swapping one coat for a similar coat is the kind of edit AI video tools handle reasonably well today.</p><p><strong>Show brands you can actually buy.</strong> Since the version can be chosen programmatically, the options don’t have to be the same everywhere. A viewer in India would see brands that ship to India, priced in rupees. A viewer in France would see European retailers in euros. A cheaper alternative isn’t much help if it can’t be delivered to you.</p><p><strong>Save or buy.</strong> From the panel, the viewer can save the look for later or add it to their bag. The MVP doesn’t need a full checkout inside the video. Handing off to the retailer’s cart is enough to test whether people use it.</p><p><strong>Learn what fans want.</strong> Every tap is recorded. Over a season, a brand partner would see which outfits got the most interest, whether viewers in a given country usually chose the original or a cheaper option, and which colours and styles people kept going back to. Right now, most of that interest disappears into screenshots and search bars.</p><h3>What the web would need</h3><p>If shopping inside video becomes common, several parts of how the internet works would need to change. <em>This is again, my view, not anything Flam has announced.</em></p><p><strong>Shopping without leaving the video.</strong> Tapping a product in a video usually sends you to a separate page. For in-scene shopping to feel natural, comparing options and adding to cart would need to happen inside the video. Apps would also need a consistent way to show what’s tappable in a moving picture.</p><p><strong>Links that open a specific version.</strong> If you want to send a friend the scene with the budget coat, the link has to open that version, not the default. Bookmarks, browser history and the back button would need to handle versions within one file.</p><p><strong>Players built for interaction.</strong> Standard web video players were built to play, pause and skip. Shoppable video needs players that handle tap areas, switching and handing off to a cart. TVs are the hardest case because people use a remote. Scanning a code with your phone while the show plays is one workaround.</p><p><strong>Delivery networks that handle variants.</strong> Content delivery networks keep copies of files close to users so they load quickly. A file with many versions inside means either storing larger files or predicting which version people in a region are likely to pick.</p><p><strong>New ad measurement.</strong> Advertising is mostly paid for by views and watch time. Taps, choices and switches would need standard definitions and independent checks before advertisers pay for them.</p><p><strong>Search that sees products inside video.</strong> Someone searching for “Emily in Paris green coat” should be able to land on the exact scene and version. That means indexing products within a video, not just its title.</p><p><strong>Clear labelling.</strong> If a scene shows Emily in a coat she never wore, viewers should be told the image was edited. Several countries are already discussing rules on disclosing AI-edited media, and it matters in fashion because people buy based on how something looks.</p><p><strong>Privacy.</strong> What someone taps says a lot about their taste and budget. Laws like India’s DPDP Act and Europe’s GDPR, and platform consent flows, would need to cover data collected inside a video.</p><p><strong>Accessibility.</strong> Tappable areas need text labels and should work with screen readers and keyboards, not only precise taps on a phone.</p><h3>What Flicks can do</h3><p><strong>Turn product pages into try-before-you-scroll experiences.</strong> Flam positions Flicks as a way to make static product pages interactive in place, letting shoppers swipe through colours, explore variants and learn about features.</p><p><strong>Make ads that capture intent, not just impressions.</strong> In digital ads, the idea is to convert scrolls into swipes and capture intent the moment someone explores. A tap on a specific variant is a far stronger signal than a view.</p><p><strong>Personalise at scale without reshooting.</strong> Swap the presenter so different audiences see someone they relate to, or swap the backdrop so a viewer in Kochi sees a monsoon street and one in Jaipur sees a desert evening.</p><p><strong>Refresh campaigns cheaply.</strong> One shoot can become a Diwali version, an Eid version and a year-end sale version by swapping scenes rather than rebooking a crew.</p><p><strong>Run creative A/B tests inside a single asset.</strong> Instead of launching three ads and splitting budget, the variants live together and viewer behaviour tells you which one wins.</p><h3>What Flicks likely can’t do today</h3><p><em>Some of these are inferences from how the format is built, so treat them as questions to test rather than confirmed limits.</em></p><p><strong>It won’t rewrite the story.</strong> It can’t change the plot, camera movements and runtime for a video. Flicks change what’s in the frame, not what happens next. If you want branching narratives, that’s a different product category.</p><p><strong>It won’t invent a variant on the spot.</strong> Because variants are generated and compressed into the asset in advance, the viewer is choosing from a set menu. Typing “show it in neon green” and getting something new instantly isn’t how the format is described.</p><p><strong>It won’t talk back.</strong> Conversation belongs to Flam’s Visual Agents, and viewing an object in 3D or placing it in your room belongs to Airboards.</p><p><strong>It may struggle with dramatic swaps.</strong> AI video editing is strongest when the replacement is similar in shape and physics. A sneaker becoming a different sneaker is easier than a sneaker becoming a knee-high boot, where hands, shadows and reflections all need to change too.</p><p><strong>It raises real consent questions around faces.</strong> Swapping people is powerful and also exactly where deepfake concerns live. Expect, and hope for, tight controls on whose likeness can be used.</p><p><strong>It hasn’t proven its lift publicly.</strong> The speed and size numbers are Flam’s own. What the market still needs is independent evidence that interactive variants actually move conversions compared with good old video.</p><h3>Where this leaves things</h3><p>Flicks isn’t just about streamlining digital ads. It is part of a broader shift in how media works. By transforming static video into an interactive canvas where viewers can explore, personalize, and choose, formats like Flicks are blurring the line between passive entertainment and active engagement. Whether it’s swapping an outfit in a show or discovering a localized product in an ad, the way we experience video content is about to get a lot more hands-on.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4cf65a51b63b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When AI video blinks: The frame consistency problem nobody talks about]]></title>
            <link>https://medium.com/@shivangibitsp/when-ai-video-blinks-the-frame-consistency-problem-nobody-talks-about-c7dd66feada0?source=rss-e0fe6055f682------2</link>
            <guid isPermaLink="false">https://medium.com/p/c7dd66feada0</guid>
            <category><![CDATA[ai-video-generation]]></category>
            <category><![CDATA[ai-video-generator]]></category>
            <category><![CDATA[video-generation]]></category>
            <dc:creator><![CDATA[SHIVANGI SINGH]]></dc:creator>
            <pubDate>Sun, 14 Jun 2026 19:38:46 GMT</pubDate>
            <atom:updated>2026-06-14T19:38:46.061Z</atom:updated>
            <content:encoded><![CDATA[<p>There are more AI video tools right now than most people can keep track of. Runway, Kling, Pika, Veo, Hailuo, Luma, Haiper, and that’s before you count the avatar tools like Synthesia and HeyGen. A new one launches every week, usually with a headline about replacing Hollywood.</p><p>This week’s Product Hunt launch, <strong>iArt.ai</strong>, is doing something different. It isn’t trying to generate realistic footage of people or landscapes. It generates motion graphics i.e. animated text, explainer videos, kinetic typography, the kind of stuff you’d see in an Apple keynote or a well-made product demo. And it does it from a text prompt in about 30 seconds.</p><p>To understand why that distinction matters, you need to understand what makes AI video generation so hard in the first place and why the market is flooded with tools that, despite looking different on the surface, are largely solving the same problem.</p><h3><strong>Why AI Video Is Not Just “Many AI Images”</strong></h3><p>Most people assume AI video works like a flipbook: generate image 1, generate image 2, generate image 3, play them fast enough and you have a video. This is exactly wrong and the reason it’s wrong explains almost every strange artifact you’ve ever seen in AI-generated video.</p><p>Think about what a camera actually captures across frames. The person’s shirt doesn’t change color between seconds 2 and 3. Their hand doesn’t grow a sixth finger and lose it again. The shadow falls in the same direction throughout the scene. All of these are <em>consistency constraints, </em>things that must stay the same across every frame because that’s how the real world works.</p><p>When AI generates each frame independently, it has no memory of what it decided in the previous frame. So a face might drift. A logo might subtly change shape. A hand might morph. This is called <strong>temporal inconsistency</strong>, and fixing it is the central engineering challenge of AI video generation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/841/1*VaN5Xr6OZJy0t1ZPbOVIgQ.png" /><figcaption>Each frame was generated without remembering what the previous one decided. This is temporal inconsistency</figcaption></figure><h3><strong>How the model actually works: Diffusion, Explained simply</strong></h3><p>The architecture underlying almost all modern AI image and video generation is called a <strong>diffusion model</strong>. Here’s the intuition:</p><p>Imagine you start with a completely blurry, noisy image i.e. pure static, like a TV with no signal. A diffusion model has learned, from looking at billions of images, how to gradually “clean up” that noise into something coherent, guided by a text prompt. It doesn’t paint an image from scratch; it refines noise into structure, step by step.</p><p>For a single image, this works beautifully. The model denoises a 2D grid of pixels (height × width) over dozens of steps until an image appears.</p><p>For video, the problem gets much harder. Now you’re denoising a 3D block, height × width × time. And every frame in that block has to be consistent with every other frame. Frame 23 needs to “remember” that frame 1 decided the character has brown hair and a red jacket.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/935/1*gL-V0Hd-rBIrE5Gvb0y2vg.png" /><figcaption>Going from image to video means adding an entire dimension of consistency constraints</figcaption></figure><h3>The technical fix: Temporal attention</h3><p>The solution the field landed on is called <strong>temporal attention</strong>. In plain terms: while the model is generating each frame, it also “looks at” the other frames and asks: “Does this frame agree with what I decided in the others?”</p><p>This cross-frame checking is computationally expensive. But it’s what prevents the shirt from changing color mid-video. It’s also what makes video generation so much slower and costlier than image generation.</p><p>There’s a second piece of engineering: before the denoising process even starts, a component called a <strong>Video VAE (Variational Autoencoder)</strong> compresses the video into a much smaller representation. A four-second video at 720p has roughly 66 million pixel values. The VAE squeezes this down by 8 to 16× in the spatial dimensions and 4 to 8× in the time dimension, giving the model something small enough to work with. When the VAE rebuilds the full video afterward, the compression itself enforces smoother transitions between frames.</p><p>Even with all of this, the model still drifts. Researchers have documented a problem called <strong>temporal drift</strong>: as the model generates frames further along in the clip, the “memory” of early frames weakens, and small deviations accumulate into visible inconsistencies. Various workarounds exist like periodically re-anchoring the model to reference frames, using optical flow (predicting where pixels should move and constraining the model to follow that trajectory), or locking attention contexts at key intervals. None of them fully solve the problem.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yPzHwCi838elxHb4ovw0Vg.png" /><figcaption>How temporal attention works?</figcaption></figure><h3><strong>Why there are so many tools and why most are doing the same thing</strong></h3><p>Here’s something the AI video market doesn’t advertise clearly: most of the major tools are solving the same core problem with the same core architecture. They differ mainly in emphasis and interface design.</p><p>Kling (from Chinese tech company Kuaishou) is optimized for human motion. It produces unusually natural body movement and facial expressions. Runway layers creative control on top of its model like motion brushes, camera angle presets, style transfer. Pika focuses on short-form social content with creative visual effects. Google Veo 3 brings Google’s infrastructure scale. But all four are fundamentally in the business of generating photorealistic footage that holds together over time. Their consistency problem is the same.</p><p>This is why, despite the crowded market, no single tool has “won.” By mid-2026, the differentiation has become narrow: Runway for high-quality brand hero content, Kling for volume at lower cost, Pika for stylized social experiments, Veo for developers who want API access. OpenAI’s Sora, which launched with enormous hype in 2024, became a cautionary tale about how expensive this problem is to solve at scale, reportedly costing $15 million per day in compute while generating a fraction of that in revenue. The Sora API was discontinued in mid-2026.</p><p>But there’s a third category that most people don’t separate out.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rWAsarcBpTrtVUnFgVso0Q.png" /><figcaption>AI video generation market map</figcaption></figure><h3>The three types of AI video (They’re not the same problem)</h3><p>Once you see this, you can’t unsee it. “AI video generation” actually describes three different technical problems:</p><p><strong>Type 1: Text-to-Video (photorealistic footage)</strong> You type “a golden retriever running through a sunlit park” and the model generates a clip that looks like real footage. The hard problems: physics (does the dog’s fur move like real fur?), character consistency (does the same dog look the same in the next scene?), and duration (most models start breaking down after 10–20 seconds). This is Runway, Kling, Veo, Sora.</p><p><strong>Type 2: Image/Avatar-to-Video (talking heads)</strong> You give the model a still photo of a face and a script, and it makes the face appear to speak. This is technically easier than Type 1 because you’re not generating from scratch; you’re warping a reference image. The face already exists; you just have to animate its mouth and expressions. This is why HeyGen and Synthesia have been production-ready for years while photorealistic footage generation still struggles. The consistency problem is simpler: keep one face coherent, not an entire scene.</p><p><strong>Type 3: Structured motion graphics</strong> This is iArt.ai’s category and the difference from Types 1 and 2 is more fundamental than it first appears.</p><p>A natural question here: can’t you just give a tool like Runway your brand hex code and logo and get consistent brand output? You can, and most tools support it. But what happens under the hood is different from what you might expect.</p><p>When you hand a photorealistic footage tool a logo or a hex color, it uses that as <strong>soft conditioning</strong>. The logo goes through the diffusion model. It’s processed as pixels, run through the same probabilistic denoising pass as everything else. The model has learned to associate your visual inputs with similar-looking outputs, but it’s still <em>predicting</em> what your brand looks like frame by frame. The result: your hex color #1A2B3C might come out as #1C2D3E two seconds later. Your logo&#39;s sharp corner might round slightly by the third scene. These aren&#39;t bugs. They&#39;re the fundamental nature of a model that generates by approximation, not by execution. The model is drawing your brand from memory, not placing it.</p><p>Motion graphics tools work differently at the asset level. Your logo is uploaded as a vector file, an SVG or Lottie element and placed directly into the composition. It is never run through a generative model. The hex color is written to every frame as an exact value, not inferred. The AI decides <em>when</em> your logo appears, <em>how</em> it moves, and <em>what timing</em> it follows, but it never touches <em>what your logo looks like</em>. That part is deterministic: it’s executed, not predicted.</p><p>The analogy: asking Runway to use your brand is like asking a very skilled artist to paint your logo from a reference photo. The result will look close. Asking iArt is like handing a choreographer your logo as a physical object and saying “make this move.” The logo itself never changes, only the motion does.</p><p>This distinction matters practically. Motion graphics tools can guarantee pixel-level brand fidelity across every frame because brand assets are never passed through the probabilistic engine. Footage tools can approximate it well and they’re getting better, but they cannot guarantee it, because approximation is built into how diffusion works.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XXi-xLUxHBkMmrBjiYQWXw.png" /><figcaption>Soft Conditioning vs Hard Placement</figcaption></figure><h3>What iArt.ai is actually building</h3><p>iArt.ai launched on Product Hunt this week as the “AI-powered After Effects alternative for non-animators.” If you’ve never opened After Effects, here’s what it means: After Effects is the professional software motion designers use to create animated videos. Setting up a single 10-second animated text sequence requires manually placing and timing every element, where each word enters, how fast it moves, what easing curve it follows, when the color shifts. For a designer, this takes 2–6 hours. The software itself has a learning curve measured in months.</p><p>iArt replaces all of that with a prompt. Type what you want, and an AI agent handles the design, animation, and timing. The company claims 30 seconds for outputs that would take hours manually.</p><p>The technical approach is meaningfully different from the photorealistic footage tools. Rather than running a single diffusion pass over a video latent, iArt’s agent builds the animation <strong>scene by scene, </strong>making layout decisions, typography choices, and timing decisions as discrete, composable steps. This is why it can maintain brand consistency across multiple outputs: an intro, a lower-third graphic, and an end card can all share the same visual language, because the agent is applying the same defined rules across each, not generating them independently from noise.</p><p>The input flexibility is also unusual. iArt accepts text prompts, Figma design files, Notion documents, images, audio files, and YouTube links. Dropping in a Figma mockup and having it animate while preserving your exact brand colors and typefaces is a fundamentally different proposition from “type a prompt, get a clip.” The former is a design tool with AI; the latter is a generative model with a UI.</p><p>The output formats signal who this is actually for: MP4, GIF, and Lottie JSON. Lottie is the file format used to embed animations directly into websites and apps. This suggests iArt isn’t only targeting marketers making social content. It’s targeting developers and product teams who need animated UI elements without keeping a motion designer on staff.</p><blockquote>Where it still struggles is what the industry calls the hard-edge consistency problem. When source assets have sharp, precise edges like a logo with crisp geometry, UI elements with pixel-perfect borders, dense typography at small sizes; frame-to-frame fidelity can break down at the points where the AI is still making decisions rather than placing fixed assets.</blockquote><p>This happens because even in a motion graphics pipeline, not everything is fully deterministic. The AI agent is still making generative decisions about how elements enter the frame, how they transition between scenes, and how they behave at the edges of their defined bounding boxes. At those decision points, the model’s probabilistic nature can produce slight inconsistencies like a logo animation that looks slightly different entering scene 2 than it did entering scene 1, or a text transition that doesn’t hold the same weight across cuts.</p><p>The current mitigation is leaning harder on style guidance and component libraries, essentially giving the agent more explicit rules to follow so it has less room to approximate. Defined inputs reliably produce more consistent outputs than ambiguous ones. But this is still an active engineering problem, not a solved one.</p><p>Interestingly, this is the same frontier the photorealistic footage tools are approaching from the opposite direction. Their hard problem is imposing enough structure on a model that wants to be probabilistic. iArt’s hard problem is the last mile of fidelity where defined constraints meet a model that still approximates at the edges. Different tools, same underlying tension between control and generation.</p><h3>Reading the market clearly</h3><p>The useful mental model for the AI video market in 2026 isn’t a ranking. It’s a segmentation by what consistency problem each tool is solving:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fUEdstB05FCrmwQlA4lI_w.png" /></figure><p>iArt.ai’s bet is that the motion graphics segment is genuinely underserved. The photorealistic footage tools don’t understand intentional typography or design layout as they generate pixels, not vector animations with brand rules. The traditional software (After Effects) requires expertise most people don’t have and time most teams don’t have. If iArt can reliably deliver brand-consistent, audio-synced motion graphics from a prompt or a Figma file, it occupies a position that none of the Runway/Kling/Pika cluster is actually competing for.</p><p>The hard-edge consistency problem is real, and the founders know it. But the market gap is real too and the positioning is sharp.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c7dd66feada0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Fundraising: not a database problem, but a graph problem]]></title>
            <link>https://medium.com/@shivangibitsp/fundraising-not-a-database-problem-but-a-graph-problem-8a608d5efa88?source=rss-e0fe6055f682------2</link>
            <guid isPermaLink="false">https://medium.com/p/8a608d5efa88</guid>
            <category><![CDATA[product-hunt]]></category>
            <category><![CDATA[fundraising]]></category>
            <category><![CDATA[bipartite-graph]]></category>
            <dc:creator><![CDATA[SHIVANGI SINGH]]></dc:creator>
            <pubDate>Sun, 07 Jun 2026 19:31:07 GMT</pubDate>
            <atom:updated>2026-06-07T19:31:07.868Z</atom:updated>
            <content:encoded><![CDATA[<p>Most founders approach fundraising like a search problem. Open PitchBook, filter by stage and sector, export a list of 200 investors, start emailing. Three months later, they’ve sent 600 follow-ups and landed 4 meetings.</p><p>The issue isn’t effort. It’s the model. <strong>Fundraising isn’t a search problem, it’s a graph problem.</strong> The right investor isn’t the one whose website says the right things. It’s the one whose last 20 deals match your profile, who is actively deploying right now, and who you can reach through someone who already has their trust.</p><p>That’s the bet <em>Fundraisly </em>is making. Built by founders who raised over $1B and an ex-investment analyst from a $600M+ AUM VC fund, it’s an AI agent that scans 300,000+ investors and millions of deals, then does the research, matching, and outreach for you. The result, they claim: 20–40 qualified investor meetings within 90 days.</p><p>Let’s look at how it actually works and what’s happening under the hood.</p><h3>The four-step engine</h3><p><em>Fundraisly’s </em>matching pipeline has four distinct layers. Each one filters and enriches the candidate list before passing it to the next.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Vj0mnlvNR4zCR3oiM6XJQw.png" /><figcaption>The Fundraisly pipeline overview</figcaption></figure><h4>1. Hard Filter: First, cut the noise</h4><p>300,000+ investors is an unusable list. Step one is a basic but critical triage: filter to only investors who actually match your hard constraints i.e. investment stage, geography, check size, and sector. This is the part every database tool does. It gets you from 300K down to a qualified pool of maybe a few hundred.</p><p>Think of it as an SQL query over a structured investor dataset:</p><pre>-- Pruning 300K investors to a qualified candidate pool<br>SELECT investor_id, fund_name, partner_name, check_size<br>FROM   investors<br>WHERE  investment_stage IN (&#39;seed&#39;, &#39;series_a&#39;)<br>  AND  geography IN (&#39;US&#39;, &#39;India&#39;)<br>  AND  sector_tags @&gt; ARRAY[&#39;fintech&#39;, &#39;b2b_saas&#39;]<br>  AND  min_check &lt;= 1000000<br>  AND  max_check &gt;= 500000<br>ORDER BY last_deal_date DESC;</pre><p>This alone isn’t enough. Most investors keep their stated thesis vague or outdated. The next step is where Fundraisly starts to diverge from a regular database tool.</p><h4>2. Behaviour Score: Ignore what investors say. Watch what they do.</h4><p>A fund’s website might say “we invest in enterprise software.” Their last 20 deals might show 60% consumer fintech. That gap is called <em>thesis drift</em> and it’s common enough that pitching against a fund’s stated thesis is genuinely risky.</p><p>Fundraisly’s second layer runs automated analysis on <em>completed deals only, </em>not the fund’s description page. For each investor in the filtered pool, it builds a behavioral profile: what sectors did they actually back in their last fund cycle? What stage, check size, and business model patterns emerge from real portfolio data? How recently have they deployed capital in your specific sub-vertical?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*INeUEjMbdb0jHfyLOLU_Tw.png" /><figcaption>What the behavioural scoring model looks like?</figcaption></figure><p>Under the hood, this is a supervised ranking model trained on historical deal data. Each investor in the pool gets a vector of behavioral features, and a learned scoring function ranks them by predicted fit to the founder’s company profile.</p><pre># Simplified behavioral scoring model<br>import numpy as np<br><br>def score_investor(investor_deals, founder_profile):<br>    &quot;&quot;&quot;<br>    investor_deals: list of past portfolio companies with metadata<br>    founder_profile: dict with sector, stage, model, geography<br>    &quot;&quot;&quot;<br>    features = {<br>        &quot;sector_overlap&quot;: compute_sector_overlap(<br>            investor_deals, founder_profile[&quot;sector&quot;]<br>        ),<br>        &quot;recency_score&quot;: deals_in_last_18_months(<br>            investor_deals, founder_profile[&quot;sector&quot;]<br>        ),<br>        &quot;stage_match&quot;: stage_alignment(<br>            investor_deals, founder_profile[&quot;stage&quot;]<br>        ),<br>        &quot;check_size_fit&quot;: check_size_alignment(<br>            investor_deals, founder_profile[&quot;raise_amount&quot;]<br>        ),<br>    }<br>    # Weighted dot product — weights learned from historical outcomes<br>    weights = np.array([0.35, 0.30, 0.20, 0.15])<br>    score   = np.dot(np.array(list(features.values())), weights)<br>    return round(score * 100, 1)</pre><p>The key insight: <em>recency is heavily weighted</em>. A fund that was active in your space two fund cycles ago but hasn’t deployed there recently is ranked down. Only current deployment signals count.</p><h4>3. Warm Path Map: The graph underneath your contacts</h4><p>This is the most technically interesting step and the one that most tools skip entirely.</p><p>A perfectly scored investor you have to cold-email is worth significantly less than a moderately good fit you can reach through someone who already knows them. Warm introductions convert at 10–20x the rate of cold outreach. The challenge is that most founders don’t know their own network well enough to find these paths.</p><p>Fundraisly connects your Gmail, Outlook, and LinkedIn to build what is effectively a <em>relationship graph, </em>a network where you are a node, investors are nodes, and edges represent real connections weighted by communication frequency, recency, and responsiveness.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xYh0GbFwqSkdqP5YNSegZQ.png" /><figcaption>Relationship graph: Finding warm paths</figcaption></figure><pre># BFS-based warm path finder over a weighted relationship graph<br>from collections import deque<br><br>def find_warm_path(graph, founder_id, target_investor_id, max_hops=3):<br>    &quot;&quot;&quot;<br>    graph: dict of {node_id: [(neighbor_id, edge_weight), ...]}<br>    Returns shortest warm path if one exists within max_hops<br>    &quot;&quot;&quot;<br>    queue   = deque([[(founder_id, 1.0)]])  # (node, cumulative_weight)<br>    visited = {founder_id}<br><br>    while queue:<br>        path = queue.popleft()<br>        current_node, _ = path[-1]<br><br>        if current_node == target_investor_id:<br>            return path<br><br>        if len(path) &gt;= max_hops:<br>            continue<br><br>        for neighbor, weight in graph.get(current_node, []):<br>            if neighbor not in visited:<br>                visited.add(neighbor)<br>                queue.append(path + [(neighbor, weight)])<br><br>    return None  # no warm path found -&gt; falls back to cold outreach</pre><p>Edge weights in this graph come from three sources: Gmail and Outlook metadata (how often do you email this person, do they reply, how recent?), and LinkedIn second-degree connections via a GDPR-certified third-party service. Fundraisly is CASA certified for its Google integration, meaning its data handling practices have been independently audited.</p><blockquote>“A perfectly matched investor you can reach through two degrees of your network is worth 10x a cold contact with identical criteria on paper.”</blockquote><h4>4. Sequenced Outreach: The agent that does the asking</h4><p>Once the ranked, warm-path-mapped investor list is ready, Fundraisly runs the actual outreach. For investors with a warm path, it drafts intro request messages. For the rest, it writes personalized cold emails; personalized against actual deal history, not just a name-swap template.</p><p>What separates this from a mass-email tool is the sequencing logic: messages are spaced, follow-ups are measured, and the sequence <em>stops the moment a reply comes in</em>. The target list is built for accuracy; investors who aren’t a genuine fit don’t make it onto the list in the first place, which is why the open rates Fundraisly reports (60–70%) run far higher than typical cold outreach averages (~27%).</p><pre># Sequencing logic - &gt; simplified state machine<br>class OutreachSequence:<br>    STEPS = [<br>        {&quot;day&quot;: 0,  &quot;type&quot;: &quot;intro_or_cold&quot;, &quot;personalized&quot;: True},<br>        {&quot;day&quot;: 5,  &quot;type&quot;: &quot;follow_up_1&quot;,   &quot;personalized&quot;: True},<br>        {&quot;day&quot;: 12, &quot;type&quot;: &quot;follow_up_2&quot;,   &quot;personalized&quot;: False},<br>    ]<br><br>    def next_action(self, investor_id, current_day):<br>        status = self.get_status(investor_id)<br><br>        # Hard stop -&gt; reply received at any point<br>        if status[&quot;replied&quot;]:<br>            return {&quot;action&quot;: &quot;stop&quot;, &quot;reason&quot;: &quot;reply_received&quot;}<br><br>        for step in self.STEPS:<br>            if current_day &lt; status[&quot;start_day&quot;] + step[&quot;day&quot;]:<br>                continue<br>            if not status[&quot;sent&quot;].get(step[&quot;type&quot;]):<br>                return {&quot;action&quot;: &quot;send&quot;, &quot;step&quot;: step}<br><br>        return {&quot;action&quot;: &quot;complete&quot;}</pre><h3>What this adds up to</h3><p>Taken together, the four steps form a pipeline that moves from a 300K+ raw universe to a personalized, warm-path-sorted shortlist of investors who are actually deploying capital in your space right now and then executes the outreach automatically.</p><p>The underlying architecture, a bipartite investor-founder graph with behavioral embeddings and shortest-path warm routing is the same class of system that recommendation engines at Netflix or Spotify are built on. The difference is that here, the stakes of a bad recommendation aren’t a forgettable movie. They’re months of a founder’s time.</p><p>If <em>Fundraisly’s</em> numbers hold at scale, the implication is significant: the network advantage that top-tier founders in major startup hubs have always had, knowing the right people, getting warm intros, reading which VCs are actually active right now, may finally be something you can run as an algorithm.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8a608d5efa88" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why your AI budget exploded and the architecture that fixes it]]></title>
            <link>https://medium.com/@shivangibitsp/why-your-ai-budget-exploded-and-the-architecture-that-fixes-it-d2a6ad94a8ae?source=rss-e0fe6055f682------2</link>
            <guid isPermaLink="false">https://medium.com/p/d2a6ad94a8ae</guid>
            <dc:creator><![CDATA[SHIVANGI SINGH]]></dc:creator>
            <pubDate>Sun, 31 May 2026 19:12:30 GMT</pubDate>
            <atom:updated>2026-05-31T19:27:04.484Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/947/1*J_Ngr65AK1VlrU1Uby_pWw.png" /><figcaption>News articles from last week</figcaption></figure><p>Microsoft cancelled Claude Code licenses for thousands of engineers in its Experiences &amp; Devices division this month. The deadline is June 30. Uber burned its entire 2026 AI coding budget in four months, by April, the money was gone. <em>The thing both companies said publicly wasn’t that the tools were bad. It was that they couldn’t afford to keep them.</em></p><p>Per-engineer API costs at Uber ran between $500 and $2,000 per month. Adoption had jumped from 32% to 84% of the engineering organization in six weeks. 70% of committed code was AI-generated. And Uber’s COO could not draw a clear line between that spend and better consumer features. Healthcare enterprises have reported consuming a trillion tokens over six months, over $6 million in unplanned costs, before finance teams understood what was driving it.</p><p>The FinOps Foundation’s 2026 report found that 73% of enterprises reported AI costs exceeding original projections. And this happened even as per-token prices fell. Average enterprise token costs dropped 67% year-over-year from $18.40 per million tokens in Q1 2025 to $6.07 in Q1 2026. Price fell. Bills rose. That contradiction is the whole story.</p><h3>The Math that breaks every budget model</h3><p>Traditional software procurement is linear. You buy seats. Usage scales roughly with headcount. Your model for Year 2 is Year 1 times a growth factor. AI on token-based billing doesn’t behave this way. It behaves like cloud compute but faster, and with a consumption curve that most finance teams have never seen before.</p><p>The mechanism is straightforward. Each agentic workflow, a debugging session, a code review, a document synthesis doesn’t make one API call. It makes dozens, sometimes hundreds. The agent reads files, generates intermediate reasoning, writes code, checks its own output, reads error messages, revises. Every one of those intermediate steps consumes tokens. At Claude Opus 4.7’s pricing of $5 per million input tokens and $25 per million output tokens, a single complex agentic loop can cost $1–5 before it terminates.</p><p>Then multiply by 5,000 engineers running that loop fifty times a day. Enterprise token consumption has grown 13x since January 2025. Volume, not price, is what broke the budgets.</p><h3>Why most enterprises are doing routing wrong?</h3><p>The instinctive response to a rising AI bill is to cut usage or negotiate pricing. Both are wrong, for the same reason: they treat a volume problem as a price problem. The architectural answer is intelligent model routing, dispatching each task to the cheapest model capable of completing it at acceptable quality.</p><p>This is harder than it sounds, and most implementations make one of three mistakes.</p><h4>Mistake 1: Using token count as a complexity signal</h4><p>The most common routing heuristic is prompt length. Long prompt = hard task = expensive model. Short prompt = easy task = cheap model. This correlation is weak enough to be dangerous. A single-sentence question — <em>“What does our India data privacy policy say about cross-border transfers?” </em>can require deep organizational knowledge and careful reasoning. A 10,000-token document can be trivially summarized by any model. Routing on length misfires in both directions, and the cost of a wrong downgrade (a subtly wrong answer that no one catches) is worse than the cost of a wrong upgrade (wasted tokens).</p><h4>Mistake 2: Stateless context injection</h4><p>Standard retrieval-augmented generation (RAG) is stateless. Every query starts from zero: embed the question, retrieve chunks from a vector store, stuff them into the context window, send. This works, but it’s expensive. You’re re-embedding, re-retrieving, and re-constructing context on every single call. For an organization with a complex internal knowledge base, that means your input token count is high by default before the model has even started reasoning.</p><p>The 2026 production architecture pattern that’s emerging from enterprise AI deployments moves through four stages: <strong>vector search identifies relevant documents and entity entry-points</strong>, <strong>graph traversal follows relationship edges to gather connected context</strong>, <strong>memory retrieval injects session continuity</strong>, and only then does the <strong>LLM generate a response</strong>. The graph layer is the key difference from standard RAG; it can answer structural questions (<em>“who owns the payments service?”</em>, <em>“what’s the approval chain for this policy exception?”</em>) through traversal rather than embedding similarity, which means the context passed to the model is smaller and more precise.</p><h4>Mistake 3: One model fits all</h4><p>Enterprises that deployed Claude Code at scale in late 2025 made an understandable mistake: they found a model that worked, and they sent everything to it. The tool had no routing, no complexity classification, no tiering. Every query whether it was auto-completing a variable name or resolving a multi-service race condition, went to the same frontier model at the same per-token price.</p><p>Production data from Q1 2026 tells a consistent story. Organizations running a tiered model architecture achieved a median blended cost of $2.31 per million tokens. Organizations routing everything to frontier models paid $18.40 per million tokens. That 87% gap is the direct consequence of one architectural decision made at deployment and never revisited.</p><h3>What good routing architecture actually looks like?</h3><p>The architecture question isn’t just which models you use. It’s how tasks move through the system. Here is how a well-designed routing stack works and why each layer matters.</p><h4>Layer 1: Context compression before the model sees anything</h4><p>The cheapest token is the one you never send. Before a query reaches any model, a well-designed system reduces the context to only what’s necessary. This is where the knowledge graph architecture matters: instead of embedding an entire document corpus and stuffing the top-k chunks into the context window, a graph-structured knowledge layer answers structural questions through traversal, following entity and relationship edges rather than semantic similarity matching.</p><p>The practical difference: a RAG system asked <em>“which teams need to sign off on cross-border data transfers?”</em> retrieves policy document chunks and hopes the answer is in them. A graph-structured system traverses the org chart, the policy ownership graph, and the approval workflow directly. The answer is shorter, more precise, and arrives with less context in the prompt, which means lower input token costs on every subsequent call.</p><h4>Layer 2: The routing decision itself</h4><p>The routing classifier should not be a heuristic. It should be a model, specifically, a fast, cheap model whose only job is to read a task, assess its complexity, and decide which tier handles it. The key design decision is the <strong>confidence threshold</strong>: what happens when the classifier is uncertain?</p><p>The right answer is: <strong>default upward</strong>. A wrong downgrade, sending a complex task to a cheap model and getting a subtly wrong answer, is more expensive in the long run than a wrong upgrade. The cost of a bad answer in production (rework, downstream errors, trust erosion) almost always exceeds the marginal token savings. Set confidence bands wide; let ambiguous tasks escalate.</p><h4>Layer 3: The feedback loop i.e. the hardest part</h4><p>Here is the problem nobody talks about in routing architecture discussions, and it’s the one that matters most in production. A wrong downgrade is usually invisible. The cheaper model returns a plausible-sounding answer. It might be subtly wrong, a summary that loses a legal nuance, a code suggestion that works in 95% of cases but breaks on an edge condition. In production, there is no ground-truth label to catch it. The savings number stays green on the dashboard right up until trust erodes and engineers stop using the system.</p><p>The only structural solution is to make routing decisions legible to users. If <strong>someone can rerun the same query against a different model tier and compare outputs side by side</strong>, two things happen: they build calibrated trust in the cheaper model when it holds up, and they surface cases where it doesn’t, which becomes training signal for refining the routing classifier. This feedback loop is the difference between a routing system that works on Day 1 and one that improves over time. Without it, routing quality can drift without anyone knowing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NPqniNdLXUosDmSWJkK1RA.png" /><figcaption>What a good routing architecture looks like?</figcaption></figure><h3>Where Coworker AI fits?</h3><p>This is the architecture context in which Coworker AI and understanding it is more useful than reading the product description. Their system implements all three layers: <strong>a hybrid graph database plus vector embedding knowledge layer (OM2) for context compression</strong>, <strong>a frontier open-weight model as the routing classifier</strong>, and <strong>a model comparison feature that exposes routing decisions to users</strong>.</p><p>The cost claim, 80% reduction, or 5x the tokens for the same spend is a blended cost-per-task figure, not a benchmark number. It holds when the majority of queries route to open-weight models like Kimi K2.6 (priced at $0.60/$2.50 per million tokens against Opus 4.7’s $5/$25), while complex tasks still reach frontier capability. On SWE-Bench Pro, the harder real-world coding benchmark, Kimi K2.6 scores 58.6% against Opus 4.7’s 64.3%. That gap exists; it’s where the routing decision earns its keep.</p><p>What’s genuinely interesting about the OM2 architecture isn’t the graph database in isolation as GraphRAG is now table stakes in enterprise AI deployments. It’s that the context compression happens <em>before</em> routing. If your prompts are shorter because the knowledge graph answered structural questions without a context dump, your input token cost is lower before you’ve even chosen a model. The two savings mechanisms stack rather than competing.</p><p>The real open question and it’s a good one to hold is whether the routing classifier is calibrated well enough at enterprise scale, and whether the feedback loop from model comparisons is tight enough to catch drift. These are hard engineering problems. They’re also the right problems to be working on in May 2026, when every CFO in a company that deployed agentic AI is asking the same question Microsoft and Uber already answered the hard way.</p><h3>What this means if you’re building or buying?</h3><p>If you’re building internal AI tooling: the single highest-leverage architectural decision you can make is introducing a routing layer. Production data suggests approximately 85% of enterprise queries can be handled by budget-tier models without quality loss, with a typical <strong>85/10/5 split (budget / balanced / frontier) yielding roughly 92% savings </strong>compared to exclusive frontier model usage. But the routing classifier needs to be a model, not a heuristic, and the confidence threshold needs to default upward.</p><p>If you’re evaluating vendor solutions: the questions to ask are not about which models are supported. They are: How does context compression work before routing? What is the confidence threshold design? How do you know, after the fact, whether a routing decision degraded output quality? What does the feedback loop look like at scale?</p><p>The enterprise AI cost crisis that made headlines this month isn’t going away. Gartner forecasts AI agent software spending at $207 billion in 2026 i.e. up 139% year-over-year and projects that agentic AI will require far more tokens per task than standard models, even as per-token prices fall. More capability means more consumption means higher bills, unless the architecture is designed to prevent it.</p><p>That’s the problem. Routing, done correctly, is the solution. Coworker AI is one implementation of that answer. There will be others. The architecture is what matters.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d2a6ad94a8ae" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[‘Dropy’ and the engineering behind price tracking extensions]]></title>
            <link>https://medium.com/@shivangibitsp/dropy-and-the-engineering-behind-price-tracking-extensions-bc54671d08ff?source=rss-e0fe6055f682------2</link>
            <guid isPermaLink="false">https://medium.com/p/bc54671d08ff</guid>
            <category><![CDATA[googleextensions]]></category>
            <category><![CDATA[product-hunt]]></category>
            <category><![CDATA[price-tracking]]></category>
            <category><![CDATA[price-tracking-tool]]></category>
            <category><![CDATA[dropy]]></category>
            <dc:creator><![CDATA[SHIVANGI SINGH]]></dc:creator>
            <pubDate>Mon, 25 May 2026 18:09:35 GMT</pubDate>
            <atom:updated>2026-05-25T18:09:35.786Z</atom:updated>
            <content:encoded><![CDATA[<p>Price drop alerts sound simple. But scraping thousands of retail sites reliably, detecting fake discounts, and delivering real-time notifications is a genuinely hard engineering problem. Here’s how it works.</p><p>Last week on Product Hunt, <em>Dropy </em>— a Chrome extension for tracking prices across Amazon, eBay, AliExpress, and thousands of other stores, caught my attention. Not because price trackers are new, but because I wanted to understand what it actually takes to build one that works at scale. Turns out there’s more going on under the hood than a simple “check the price every hour” script.</p><h3>The core architecture problem</h3><p>A price tracking extension has to solve three distinct, non-trivial engineering problems simultaneously:</p><ol><li><strong>Data extraction</strong> — scraping structured price data from wildly inconsistent HTML across thousands of websites.</li><li><strong>Change detection</strong> — efficiently knowing when a price has changed without hammering every site continuously.</li><li><strong>Delivery </strong>— pushing notifications reliably to a user in near-real time across sessions.</li></ol><p>Each of these has interesting constraints. Let’s unpack them.</p><h3>Layer 1: How price data is extracted</h3><p>Modern e-commerce sites are not friendly to scrapers. They use dynamically rendered JavaScript (React, Next.js, Angular storefronts), bot detection middleware, and anti-scraping CAPTCHAs. A naive approach, fetching raw HTML and parsing it, fails immediately on most major retail pages.</p><p>Tools like <em>Dropy </em>almost certainly use a combination of approaches:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/577/1*K6C8SWrHi75AEjbMipG3yw.png" /></figure><p>The cleanest path is JSON-LD structured data. Google mandated schema.org markup for rich snippets, so many retail sites embed product price as machine-readable JSON directly in the page &lt;head&gt;. <em>Dropy&#39;s</em> extension can read this from the DOM at page-load time without a backend round-trip.</p><pre>// Runs inside the page context as a Chrome content script<br>function extractPrice() {<br>  const scripts = document.querySelectorAll(&#39;script[type=&quot;application/ld+json&quot;]&#39;);<br><br>  for (const s of scripts) {<br>    try {<br>      const data = JSON.parse(s.textContent);<br>      const offer = data?.offers ?? data?.Offer;<br><br>      if (offer?.price) {<br>        return {<br>          price: parseFloat(offer.price),<br>          currency: offer.priceCurrency ?? &#39;USD&#39;,<br>          availability: offer.availability<br>        };<br>      }<br>    } catch (_) { continue; }<br>  }<br><br>  return null; // fall through to DOM selector tier<br>}</pre><blockquote>Imagine you walk into a store and want to know the price of something. You could either look at the price tag directly, or ask a store employee. JSON-LD is like a hidden “price tag” that websites stick in their code specifically for machines to read. Google uses it for search results, so most big retailers already have it.</blockquote><blockquote>This snippet is the extension’s eyes. When you land on a product page, it quietly reads that hidden price tag in the background. If it finds one, it grabs the number and the currency. If it doesn’t find one (not every site has it), it shrugs and says “okay, I’ll try something else” — that’s the return null at the end, which means &quot;fall through to the next method.&quot;</blockquote><p>When JSON-LD isn’t available, the extension falls back to a site-specific CSS selector ruleset — essentially a curated database of .price-now or #priceblock_ourprice selectors maintained per domain. This is brittle but fast. The really interesting engineering is in Tier 4: using ML to generalize across unknown sites by identifying price-like text nodes near &quot;add to cart&quot; buttons.</p><h3>Layer 2 : The backend polling problem</h3><p>Once a user adds an item to their watchlist, <em>Dropy’s</em> backend has to keep checking that URL for price changes. This sounds simple. It isn’t.</p><blockquote>“If 100,000 users are tracking 5 items each, that’s 500,000 URLs your system has to check repeatedly, forever, without getting rate-limited or blocked.”</blockquote><p>The smart approach here is deduplication. If 4,000 users are all tracking the same AirPods listing on Amazon, you only need to check that URL once per polling interval, not 4,000 times. The backend maintains a URL → [user_ids] map, and a single scrape result is fanned out to all interested subscribers.</p><p><strong>Adaptive polling frequency</strong></p><p>Checking every item every minute would be both expensive and easily blocked. A well-designed system uses adaptive polling, items with a history of frequent price changes are checked more often; stable items less so. This is essentially a lightweight time-series volatility model on each product’s price history.</p><blockquote>Engineering Insight: Amazon’s product pages are heavily bot-protected. Tools like Dropy likely use a combination of residential proxy rotation, request throttling, and browser fingerprint randomization (via Puppeteer Stealth or similar) to avoid detection. Some services use Amazon’s official Product Advertising API for 1st party data — much more relaiable, but restricted to affiliate partners.</blockquote><h3>Layer 3: Detecting fake discounts (the real moat)</h3><p>This is where <em>Dropy’s </em>value proposition goes beyond simple scraping. Retailers routinely inflate the “original price” before a sale to make discounts look larger. Amazon’s own research has shown that many “Lightning Deal” prices were not actually lower than the item’s average price in the preceding 30 days.</p><p><em>Dropy</em> stores a rolling time-series of every price it has observed for a product. When displaying a “price drop,” it can compute whether the current price is genuinely below the 90-day moving average, not just below an inflated reference price set by the retailer.</p><pre>function isGenuineDrop(history, currentPrice) {<br>  const last90 = history.filter(p =&gt;<br>    p.timestamp &gt; Date.now() - 90 * 24 * 3600 * 1000<br>  );<br><br>  const avg = last90.reduce((s, p) =&gt; s + p.price, 0) / last90.length;<br>  const allTimeMin = Math.min(...history.map(p =&gt; p.price));<br><br>  return {<br>    belowAverage: currentPrice &lt; avg,<br>    nearAllTimeMin: currentPrice &lt;= allTimeMin * 1.05,<br>    dropPercent: (((avg - currentPrice) / avg) * 100).toFixed(1)<br>  };<br>}</pre><blockquote>This answers the question: <em>“Is this price drop genuine, or did the retailer just inflate the price last week to make today’s discount look good?”</em></blockquote><blockquote>It looks at the price history of the last 90 days, calculates the average price over that period, and then compares today’s price against it. It also checks if today’s price is close to the lowest price ever recorded. So instead of trusting the retailer’s own “was $99, now $59!” label, Dropy does its own maths and tells you independently whether this is actually cheap or just dressed up to look cheap.</blockquote><p>The price history chart <em>Dropy </em>exposes to users is a direct byproduct of this data collection. It’s not just a visual nicety, it’s the output of the same time-series store the backend uses for deal verification.</p><h3>Layer 4: Notification delivery</h3><p>Chrome extensions have a lifecycle problem: service workers (the replacement for background pages) are terminated after ~30 seconds of inactivity. An extension cannot maintain a persistent WebSocket connection to receive server-push events. The browser simply won’t allow it.</p><p>The standard workaround is to use the Chrome Push API via Firebase Cloud Messaging (FCM). The extension registers a push subscription, and Dropy’s backend sends a Web Push notification through FCM’s infrastructure when a tracked price drops. The browser wakes the service worker, which handles the notification even if Chrome isn’t actively open.</p><h3>If you were building this yourself</h3><p>A minimal version of <em>Dropy</em> is a realistic weekend project for a developer. The stack would be: a Chrome content script to extract price from the current page, a lightweight backend (Node.js + cron) to poll stored URLs, a SQLite or Postgres table for price history, and a Telegram or email webhook for notifications. The hard parts like — handling JS-rendered sites, scaling deduplication, and bot evasion, are what separate a hobby tool from a production service.</p><p>What makes <em>Dropy </em>interesting as a product isn’t any single clever trick. It’s the unglamorous work of maintaining selector rulesets across thousands of constantly-changing retail sites, keeping proxy infrastructure healthy, and tuning polling frequency to stay within rate limits reliably, at scale, for a consumer audience that just wants a notification when something goes on sale.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*h4rGq8TPNAM34jwONV3e7w.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bc54671d08ff" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why evals are failing you? — Failures hide in the 99% data sampled out]]></title>
            <link>https://medium.com/@shivangibitsp/why-evals-are-failing-you-failures-hide-in-the-99-data-sampled-out-9ddc057e5666?source=rss-e0fe6055f682------2</link>
            <guid isPermaLink="false">https://medium.com/p/9ddc057e5666</guid>
            <category><![CDATA[eval]]></category>
            <category><![CDATA[evaluation]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[llm-evaluation]]></category>
            <dc:creator><![CDATA[SHIVANGI SINGH]]></dc:creator>
            <pubDate>Sat, 02 May 2026 19:15:14 GMT</pubDate>
            <atom:updated>2026-05-02T19:36:09.632Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>The hidden crisis of AI evaluation and how a new approach called “vibe training” is changing everything about how we know if our AI is actually working.</blockquote><p>Imagine you run a hospital. Every morning, a doctor reviews 1 in every 100 patient files; not all of them, just a sample and signs off that care standards are being met. The other 99? They go unreviewed. Missed prescriptions, wrong dosages, miscommunications, they happen quietly, invisibly, in the gaps between the samples.</p><p>You wouldn’t accept that in a hospital. But right now, that’s exactly how most AI teams run quality control on their AI systems.</p><p>This week’s Product Hunt #1 trending product is trying to fix that. But before we get to the solution, let’s understand the actual problem, because it’s deeper than most people realize.</p><h3>What are “Evals” and why should you care?</h3><p>When you build or deploy an AI agent — a chatbot, a code assistant, a customer service bot, you need a way to check whether it’s behaving correctly. Is it staying on topic? Is it being rude? Is it making things up? Is it following the rules you set?</p><p>This process of testing and monitoring AI behavior is called evaluation, or “evals” in the industry. Think of it as the quality control department for AI.</p><p>Now here’s the catch: who (or what) do you hire to run this quality control?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kY_lwwbrVW4W5yXfKU0xJw.png" /><figcaption>% of AI teams using each approach (2025 industry survey)</figcaption></figure><p>Most teams today use a big, powerful AI model, like GPT-4 or Claude, as the “judge.” You feed it the AI’s output and ask: “Was this response good?” The judge says yes or no.</p><p>This approach is called LLM-as-judge. And on paper, it sounds reasonable. But in practice, it’s quietly rotting your entire quality pipeline.</p><h3>The three silent killers of LLM-as-Judge</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yMxQDBSCO4y2W1tLltYYmQ.png" /><figcaption>3 silent killers of LLM-as-judge</figcaption></figure><blockquote>The failures that matter live in the tail. Sampling-based quality checks give you a false sense of security.</blockquote><p>When an AI agent fails silently gives a customer wrong information, violates a compliance rule, says something harmful businesses face real consequences: eroded trust, regulatory exposure, brand damage. And yet most teams are flying blind on the vast majority of their AI’s output.</p><h3>Enter <em>Plurai</em>: “Vibe training” for AI reliability</h3><p>This week’s most interesting product I found on Product Hunt is Plurai, a platform that’s coined a genuinely new concept: <em>vibe training</em>. If you’ve heard of “vibe coding”, where you describe what you want in plain English and AI writes the code — vibe training is the same idea but for AI quality control.</p><p>You describe what your AI agent should and shouldn’t do. Plurai figures out the rest.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1cw1Ycn_E-Ek2c5gcevOow.png" /><figcaption>Plurai’s pipeline: From description to always-on eval</figcaption></figure><h3>The “Multi-agent debate” — How clean training data gets made</h3><p>The most technically clever part of Plurai is how it creates training data. Instead of asking a single AI judge “is this good or bad?”, Plurai runs a debate between multiple AI agents; some acting as judges, others as devil’s advocates.</p><p>Only the examples that reach consensus across all agents make it into the training set. The contested ones get refined or thrown out. The result is a far cleaner, higher-quality dataset than you’d ever get from a single LLM judge, which means the small model trained on it is far more accurate.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e0ReoEXG44vKsv-uHF0KMw.png" /><figcaption>Multi-agent debate framework</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/913/1*33g5xbfKMK15MEB9qXTJvg.png" /></figure><h3>The “Always on, not sampled” shift</h3><p>The most profound change Plurai enables isn’t the cost savings or even the accuracy improvement. It’s the philosophical shift: you no longer have to choose what to check.</p><p>When evaluating is cheap enough to run on every single interaction, sub-100ms, trained small models rather than API-heavy giants — the sampling problem evaporates. You know about failures as they happen, not after they’ve already affected thousands of users.</p><h3>Why it’s trending and why it matters</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/910/1*b-L2_PY5w80RjtcdpTyh3w.png" /><figcaption>How Plurai compares against LLM judges?</figcaption></figure><h3>The bottom line</h3><p>We’re at an inflection point with AI deployment. The first wave was about building AI agents. The second wave, the one we’re entering now is about making those agents <em>reliably trustworthy at scale</em>.</p><p>Plurai is a serious bet on the infrastructure layer that makes reliability possible: fast, cheap, always-on evaluation that you can set up in minutes without any ML expertise. The fact that it’s backed by published research and produces measurably better results than the dominant approach makes it worth paying close attention to.</p><p>If you’re building with AI and especially if you’re deploying AI agents in production, the question isn’t whether you need better evals. It’s whether you can afford not to have them running on everything.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9ddc057e5666" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Which LLM should you actually use? A no-nonsense guide to picking the right model]]></title>
            <link>https://medium.com/@shivangibitsp/which-llm-should-you-actually-use-a-no-nonsense-guide-to-picking-the-right-model-32b42ee6bcde?source=rss-e0fe6055f682------2</link>
            <guid isPermaLink="false">https://medium.com/p/32b42ee6bcde</guid>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[llm-evaluation]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[openai]]></category>
            <dc:creator><![CDATA[SHIVANGI SINGH]]></dc:creator>
            <pubDate>Sun, 26 Apr 2026 18:29:50 GMT</pubDate>
            <atom:updated>2026-04-26T18:54:49.232Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>With dozens of AI models competing for your attention, the real question isn’t “which is best” — it’s “best for what?” Here’s how to think about it.</blockquote><p>Every week, a new model drops. GPT-5, Claude Opus, Gemini Ultra, Llama 4, DeepSeek V3 — the list keeps growing. And every week, someone asks the same question: “But which one should I use?”</p><p>The honest answer? There’s no single best LLM. There’s only the best model <em>for your task</em>, your budget, and your tolerance for trade-offs. Welcome to the first edition of my Product Hunt Series — a weekly breakdown of the most interesting products I find on the platform. This week, I want to solve the LLM selection problem once and for all — and introduce a tool I stumbled upon that actually makes this tractable.</p><h3>6 Parameters That Actually Matter</h3><p>Before you can compare models, you need to know what you’re comparing. Benchmark scores are a starting point — but production performance depends on these six dimensions:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Dw1qr4uMi8nDfzEzf7g0Rw.png" /></figure><h3>Popular LLMs at a Glance</h3><p>Grounding your model choice in real data rather than vendor claims is the fastest way to cut through the noise. Two sources I keep coming back to: Artificial Analysis, which tracks raw intelligence, speed, and cost across models — and Chatbot Arena, where millions of humans vote head-to-head on actual outputs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-dxMNOH6AlGTYfNANFPNhQ.png" /><figcaption><em>Comparison of LLMs; Source</em>: Artificial Analysis</figcaption></figure><blockquote>The trade-off is immediately visible: Most intelligent models often have lower speed and higher price</blockquote><p>Arena Elo is different from benchmark scores — it’s based on millions of real blind head-to-head votes by humans, making it arguably the most honest signal of real-world quality. The top contenders as of April 2026 are:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c1w-c73Jy7uMZABWUyGKKA.png" /><figcaption>LLM Ranking; Source: Chatbot Arena</figcaption></figure><h3>Match your task to a model</h3><p>Here’s a practical cheat-sheet for the LLM selection dilemma. For each model, I’ve tried to answer the question a practitioner actually asks: <em>when should I reach for this one?</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IMXx2et_PVYdjCCP78Q9HQ.png" /><figcaption>Strengths &amp; Weaknesses of some popular LLMs</figcaption></figure><p>Think of model selection as a routing problem. Below is a simplified mental model for which direction to look first, depending on your use case:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G_c6SF8I_Pc7NoOMkZvSKg.png" /><figcaption>Decision Mind Map</figcaption></figure><p>This is of course a simplification — real selection involves blending these factors. Which is exactly why manual comparison is tedious and error-prone. That brings me to this week’s featured product.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cLLi64NYrSRCN9wG7qpSYA.png" /><figcaption>About QuickCompare by Trismik</figcaption></figure><p>MMLU, AIME, Arena scores — are useful filters for <em>narrowing</em> candidates. But they have a fundamental problem: they test models on generic tasks, under ideal conditions, with standardized prompts. Your production environment is none of those things.</p><p>Consider: a model that dominates coding benchmarks may struggle with structured data extraction from your internal documents. A model scoring 95% on MMLU may hallucinate confidently when queried about your proprietary product catalogue. This is why the mantra of tools like QuickCompare matters — evaluate on your data, not someone else’s.</p><h3>The Bottom Line</h3><p>There’s no universal best LLM in 2026. GPT-5 leads on breadth. Claude Sonnet wins on long context and agentic reliability. DeepSeek R1 punches far above its cost. Llama 4 Scout gives you massive context with full data ownership. Gemini 2.5 Pro is the pick if you’re deep in Google’s orbit.</p><p>The right move is to pick two or three candidates based on your task profile, run them against a representative slice of your actual workload, and let the data decide. QuickCompare is built exactly for this workflow — and it’s the most practical tool I’ve seen this week on Product Hunt for solving this specific problem.</p><p>Next week, I’ll be back with another Product Hunt find. If you found this useful, follow along — there’s a lot more where this came from.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=32b42ee6bcde" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Are you ‘Hooked’? Why our favorite utility app is a masterclass in behavioral design]]></title>
            <link>https://medium.com/@shivangibitsp/are-you-hooked-why-our-favorite-utility-app-is-a-masterclass-in-behavioral-design-e174b6473bcc?source=rss-e0fe6055f682------2</link>
            <guid isPermaLink="false">https://medium.com/p/e174b6473bcc</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[screentime]]></category>
            <category><![CDATA[mobile-apps]]></category>
            <category><![CDATA[whatsapp]]></category>
            <dc:creator><![CDATA[SHIVANGI SINGH]]></dc:creator>
            <pubDate>Mon, 02 Feb 2026 21:23:38 GMT</pubDate>
            <atom:updated>2026-02-03T05:15:06.339Z</atom:updated>
            <content:encoded><![CDATA[<p>Last week, my phone told me something I didn’t ask to know. My average screen time was <strong>5 hours 46 minutes a day</strong>. It has been long time since I deactivated my Instagram account so no mindless scrolling. I wondered, where did all this time go?</p><p>It was <strong>WhatsApp</strong>.</p><p>That surprised me-because I’ve never once opened WhatsApp thinking,<br> <em>“Ah yes, let me enjoy some quality WhatsApp time.”</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1003/1*Z3EwllYbQy_3GJOK5e7dlA.png" /></figure><p>If I’m not using WhatsApp for pleasure…why does it still have me ‘<strong>Hooked</strong>’? Can I even think of switching to a new messaging app like Telegram, Teams or Messages?</p><p>While reading <em>Hooked</em> by Nir Eyal, I realized something uncomfortable: WhatsApp might be one of the most perfectly hooked products of our time.<br>Not because it demands attention, but because it disguises itself as responsibility.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/963/1*rDcumwaxm5bGydTLT5yWKQ.png" /><figcaption>The Hook Loop for WhatsApp</figcaption></figure><p>Here’s how WhatsApp does it — without looking like it’s trying:</p><ul><li><strong>The nudge you can’t ignore:</strong> Every buzz, number against unread chats, or typing bubble isn’t flashy — it’s guilt wrapped in social obligation. You don’t open WhatsApp for fun; you open it because you <em>feel</em> like you have to.</li><li><strong>The effortless loop:</strong> One tap, a scroll, a reply. It’s frictionless. Your brain barely needs to decide. That’s why small habits turn into hours.</li><li><strong>The unpredictable pull:</strong> Sometimes you get a reply immediately, sometimes nothing. Sometimes a meme, sometimes work chaos. That uncertainty keeps you coming back as humans are wired to check repetitively when outcomes are unpredictable — the ultimate ‘just one more time’ trap.</li><li><strong>The sunk cost trap:</strong> Every message sent, every group joined, every chat archived becomes part of your personal ecosystem. Leave the platform, and it’s not just uninstalling an app — it’s breaking habits, abandoning social threads, and betraying the version of yourself you’ve built.</li></ul><p>WhatsApp’s biggest lock-in isn’t its features.<br>It’s people.</p><p>And that’s why leaving WhatsApp doesn’t feel like uninstalling an app.<br>It feels like socially disappearing.</p><p>From a PM perspective, it’s brilliant: a product that uses <strong>internal triggers, frictionless action, variable rewards, and investment loops</strong> without ever feeling manipulative. And that’s the trick — it doesn’t fight your attention. It becomes infrastructure.</p><p>Does this mean WhatsApp’s creators are highly skilled at influencing user behavior, and does it imply a level of social irresponsibility?</p><p>Before answering that let’s look at some statistics:</p><blockquote><strong>53.1%</strong> of students felt their concentration reduced due to usage, and <strong>34.3%</strong> believed it lowered their academic scores.</blockquote><blockquote>Among specialists, <strong>36%</strong> reported significant negative effects on their married life, including reduced quality time and frequent altercations.</blockquote><blockquote>Overuse of WhatsApp is linked to vision strain (<strong>43.6%</strong>) and reduced physical activity (<strong>38.4%</strong>).</blockquote><p>These numbers paint a striking picture: WhatsApp is more than just a messaging app — it’s a force that shapes daily behavior, sometimes at the cost of concentration, relationships, or even physical health. But knowing its impact leads to a deeper question: <strong>what does this say about the ethics of its design?</strong></p><p>To answer that, we can turn to Nir Eyal’s <em>Manipulation Matrix</em> from <em>Hooked</em>, a framework that asks two simple yet powerful questions: <em>Do the creators use the product themselves?</em> and <em>Does it genuinely improve users’ lives?</em> Depending on the answers, products fall into one of four quadrants — <strong>Facilitator, Entertainer, Peddler, or Dealer — </strong>allowing us to map WhatsApp and other addictive products on a moral spectrum.</p><p>Below I have tried mapping all the apps that have hijacked my screen time… and honestly, I’m shocked yet again. The surprising insight: these apps aren’t manipulating us — they’re tools operating as designed and that would mean, screen time isn’t a failure of the platform, it’s the outcome of habit loops we’ve unconsciously reinforced. Reducing usage, then, isn’t about unistalling apps — it’s about restructuring triggers, routines, and attention patterns to align with intentional behavior.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/910/1*_B6drMZVgDTnIdN6j4gLpA.png" /><figcaption>The Manipulation Matrix for everyday mobile apps</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e174b6473bcc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Retention before acquisition: How best-in-class products reduce churn?]]></title>
            <link>https://medium.com/@shivangibitsp/retention-before-acquisition-how-best-in-class-products-reduce-churn-3c54be064913?source=rss-e0fe6055f682------2</link>
            <guid isPermaLink="false">https://medium.com/p/3c54be064913</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[retention-strategies]]></category>
            <category><![CDATA[duolingo]]></category>
            <category><![CDATA[spotify]]></category>
            <category><![CDATA[gmail]]></category>
            <dc:creator><![CDATA[SHIVANGI SINGH]]></dc:creator>
            <pubDate>Sun, 16 Feb 2025 19:16:15 GMT</pubDate>
            <atom:updated>2025-02-16T19:57:22.392Z</atom:updated>
            <content:encoded><![CDATA[<p>I often find myself, complaining to my friends, about an annoying little green owl <em>(you might have already guessed, who I am talking about) </em>who keeps on bugging me everyday with reminders to complete my French lesson, throwing tantrums and being overly dramatic when you don’t heed to its requests — or warnings. If I don’t pass a milestone, it gets upset like a parent/teacher whose ward under-performs and if I surpass its expectations I am rewarded. One day, it forced me to buy its ‘Super’ feature which is a subscription package, which allows one to make more mistakes and learn from it. Before that I would be allowed only 3 mistakes to complete my goals which I could never, making me frustrated and ultimately make this impulsive purchase from where I never turned back. And yesterday weirdly, I found myself planning a celebration with my friend when we complete our 1 year streak. <em>(We’re currently only at 150). </em>But the point is that I am motivated to open that app for another almost 2/3rd of the year.</p><p>If by chance you haven’t guessed it already, I am talking about Duolingo and it seems to have cracked customer retention and is thriving.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Yef499ZoMHRCTIEtMDqFPQ.png" /></figure><p>The longer a company can retain a customer, the greater is its Customer Lifetime Value (CLV).</p><blockquote><strong>CLV = Average transaction size * Number of transactions * Retention period</strong></blockquote><p>The retained customers not only bring revenue to the company but also reduces the acquisition costs of the customers by becoming advocates for the product. And thereby increasing the profitability of the company.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HXWMoOW3zctQRiYR9wRXwg.png" /></figure><p>In the customer lifecycle, the <strong>trial phase</strong> is where the magic happens — it’s the moment your product must prove its worth. If users don’t experience tangible value during their first-hand interaction, they’re unlikely to convert into paying customers. But value realization doesn’t stop there. It’s a continuous process that extends through retention, rewards, and advocacy.</p><h4><em>1. Pricing: The tightrope walk</em></h4><p>Pricing plays a pivotal role in this journey. Set it too high, and you risk alienating customers who feel the cost outweighs the value. Set it too low, and you inadvertently devalue your offering, making it harder to justify its worth later. The sweet spot? Pricing that aligns perfectly with the perceived value — neither overpromising nor underselling.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nFmP3Fz4_PfOwhqtu82MoQ.png" /><figcaption>News articles and social media uproar on Spotify India making key free features paid</figcaption></figure><p>The best example of this is when Spotify India in October 2023, announced paywalls for basic features that earlier used to be free like shuffle songs, add to queue, rewind songs and playlist controls, it faced a severe back lash from the users. Most of the users migrated to other free platforms like Wynk, Gaana, Jio Saavn, Sound Cloud, etc. Others who stayed without moving to premium segment was because of ‘Smart Shuffle’ feature which plays the next song based on the vibe of the current song.</p><p>The app rating which earlier used to be 4 stars, now fell to mere 2.4 stars showing the discontent of the users. Spotify’s premium plan (₹119/month) felt steep compared to local alternatives like JioSaavn (₹99/month) or Airtel Wynk (free with mobile plans). Simultaneously, local rivals started capitalizing on the restrictions imposed on the free users by Spotify-</p><ul><li><strong>YouTube Music</strong>: Offered free shuffle, rewinding, and playlist controls with ads.</li><li><strong>JioSaavn</strong>: Bundled free music with Jio’s ₹399/month mobile plan, leveraging telecom partnerships.</li><li><strong>Gaana</strong>: Allowed free users to skip ads after 5 seconds, creating a smoother experience.</li></ul><p>Spotify was hence, forced to quickly adapt for which it took following measure:</p><ul><li><strong>Cheaper Mobile-Only Plan</strong>: Launched a ₹59/month tier for cost-sensitive users.</li><li><strong>Localized Free Tier</strong>: Added Indian podcasts, devotional playlists, and Bollywood exclusives to retain free users.</li><li><strong>Partnerships</strong>: Collaborated with Airtel and Jio to bundle subscriptions with data plans.</li><li><strong>UI Tweaks</strong>: Reduced restrictions on skips/rewinds after backlash.</li></ul><h4>2. Continuous value delivery</h4><p>Once a customer pays, the real work begins. Your product must consistently deliver on its promise, helping users <strong>continuously realize the value</strong> they’re paying for. This could mean regular updates, personalized experiences, or proactive support. The goal is to make the product indispensable.</p><p>At DuoCon 2021, Duolingo introduced <em>Birdbrain</em>, the AI-powered personalization model that enhances the learning experience. Birdbrain selects exercises tailored to each user by analyzing both the difficulty of the exercises and the learner’s proficiency level. Additionally, it optimizes notification timing and messaging to keep users engaged and motivated. It is powered not only by the learning history of the user but all the users on the app.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3LL6xVPQtF8A7JslYqfh6g.png" /><figcaption>How birdbrain helps in personalization for Duolingo users?</figcaption></figure><p>Along with personalization, Duolingo continuously updates its offerings and tests them through A/B experiments. A feature you love (or dislike) today might disappear tomorrow — only to return later in a more refined form. However, retention alone isn’t enough if long streaks don’t translate into real language proficiency. If users stay engaged but fail to effectively learn and use the language, they may eventually churn. Striking the right balance between gamification and meaningful learning is key to keeping users both motivated and truly progressing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rsCcGHnxmM1puKhZiH65bA.png" /><figcaption>Regular feature updates through A/B testing</figcaption></figure><h4>3. Rewards and exit barriers</h4><p>To lock in loyalty, introduce <strong>reward mechanisms</strong> — exclusive features, discounts, or recognition for long-term users. At the same time, make exit costlier over time. This isn’t about trapping customers but about deepening their investment in your product, whether through data, habits, or emotional connections.</p><p>I recently realized I had run out of space on my Gmail account, forcing me to make a decision — should I abandon this account or pay for extra storage? The answer was obvious. Leaving would mean losing access to thousands of important contacts, migrating countless emails to another platform, and disrupting the seamless experience of Google’s ecosystem. Gmail’s intuitive UI, smart features, and deep integrations with tools like Drive, Calendar, and Meet have made managing digital communication effortless. The inconvenience of switching far outweighed the cost of additional storage.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rzbCwwTsAw6bzEGSQl6aQg.png" /><figcaption>Paying for extra storage on Gmail is value for money</figcaption></figure><p>Retention isn’t just a metric — it’s the lifeblood of sustainable growth. As we’ve seen with Spotify, Duolingo, and Gmail, the best products create experiences so valuable that leaving feels unthinkable. Whether through personalization, habit formation, or ecosystem lock-in, these brands make their services indispensable. Companies like Netflix and Apple follow the same playbook, proving that the key to long-term success isn’t just attracting users — it’s making them stay, engaged, and coming back for more.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3c54be064913" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The anatomy of Dunzo: What makes it tick?]]></title>
            <link>https://medium.com/@shivangibitsp/the-anatomy-of-dunzo-what-makes-it-tick-ebd88dfc5777?source=rss-e0fe6055f682------2</link>
            <guid isPermaLink="false">https://medium.com/p/ebd88dfc5777</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[delivery-app]]></category>
            <category><![CDATA[product-teardown]]></category>
            <category><![CDATA[product-improvements]]></category>
            <category><![CDATA[product-design]]></category>
            <dc:creator><![CDATA[SHIVANGI SINGH]]></dc:creator>
            <pubDate>Thu, 09 Jan 2025 17:06:31 GMT</pubDate>
            <atom:updated>2025-01-09T17:06:31.582Z</atom:updated>
            <content:encoded><![CDATA[<p>One of India’s earliest hyperlocal delivery service Dunzo, which kicked off its operations as early as 2014 is losing to Zomato owned Blinkit, Swiggy owned Swiggy Instamart, and Zepto. Considering the package delivery offering, Uber and Swiggy Genie also pose a strong competition. <em>(This is my personal observation based on preference of my kith and kins)</em>.</p><p>In this article I have tried to analyze the Dunzo mobile app and presented my observations highlighting the strengths and identifying the gaps that need to be addressed. Each feature has been classified as either ‘Must have’, ‘Should have’, ‘Could have’ or ‘Won’t have’ based on the <strong>MoSCoW method</strong>.</p><p>The pain points in focus here are the ones impacting the user flow and app navigation starting from product discovery <em>(or service discovery in case of package drop)</em>, add to cart, checkout, payment, delivery tracking and feedback.</p><p><strong>User persona: </strong>The user in this analysis is a Techie (20–40 yrs old) living in Gurugram in a rented flat or PG and uses the app for ordering groceries and sending/receiving small parcels <em>(ex. the house key that the user left behind at a friend’s place)</em>.</p><p><strong>Use case: </strong>The user has a craving of Yogurt and wants to order it to their home instead of going to the shop to get it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BrWs2TuDjEHqVunHbB8VdA.png" /><figcaption><strong><em>Fig. 1. Current user flow for the Dunzo app</em></strong></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/677/1*gFS_qdqqpEOkpwkEgP6mJQ.png" /><figcaption><strong>Fig. 2. Landing page after logging in and address selection</strong></figcaption></figure><p>One of the most crucial steps in the user journey of any delivery app is ensuring users can easily select their address. The Dunzo app gives three options to set the user address:</p><p>a. Search bar to add a new location</p><p>b. Use current location</p><p>c. Select from any saved location</p><p>While options <em>b</em> and <em>c</em> are one-click actions, it has been observed that when a user enters a detailed address, the search often returns <em>‘No search results found’ (as shown in screen 2, Fig. 2). </em>As a result, the user is forced to enter the area name instead of the full address to proceed, which might not be intuitive and can lead to user exiting the app at this very step.</p><blockquote><strong><em>Recommendation 1:</em> </strong>A smarter location search algorithm shall be used such that the app suggests the places closest to the user search instead of ‘No search results found’ message. [Must have and need improvement]</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1005/1*DiC6QdootZwYRnkt1qcwcA.png" /><figcaption><strong>Fig. 3. Product search flow</strong></figcaption></figure><p>In the Fig. 3. above, the screen labelled as 1 is the ‘Home Screen’, which the user lands upon after selecting the address, which is fairly simple and clutter-free which facilitates the ease of discovery <em>[Should have and done right].</em></p><blockquote>‘Groceries &amp; Essentials’ and ‘Fruits &amp; Vegetables’ are overlapping categories which might confuse the user, hence can be merged into one category.</blockquote><p><strong>Pain Point 2:</strong> The user might have multiple items within mixed categories to buy, in which case the category based store might confuse the user.</p><blockquote><strong><em>Recommendation 2: </em></strong>Adding ‘Dunzo list’ banner on the main page below the store categories can help solve the above problem. [Must have and missing]</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/322/1*KVRGfr4-eRLkb0R55tpizQ.png" /><figcaption><strong>Fig. 4. Recommended home screen</strong></figcaption></figure><p>The flow for ordering by making a list is currently as shown in Fig. 5.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/972/1*doKE5IRoWRiBcMYCGMAv6A.png" /><figcaption><strong>Fig. 5. Ordering using list — user flow</strong></figcaption></figure><p><strong>Pain Point 3: </strong>After the user selects category, enters pickup and delivery address, the pop-up for ‘Category unserviceable’ is seen which might make users feel that their effort in refilling the delivery address was in vain.</p><blockquote><strong><em>Recommendation 3: </em></strong>The delivery address could be pre-filled and only the categories that could be serviced, should be given as option for selection. If none of the options are serviceable, a banner should be displayed at top with clear message that says — ‘Current location is currently unserviceable using list’. However the option to modify the delivery address should be given to the user.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/877/1*pKALaLLbDqU80EZLHfuvdQ.png" /><figcaption><strong>Fig. 6. Recommended screens for ordering using lists</strong></figcaption></figure><p>Since, the user wants to get Yogurt they go to ‘Grocery &amp; Essentials’ store and they land on the screen 2 as shown in Fig. 3. Here, 143 nearby stores are listed, sorted in increasing order of their distance <em>(not the expected delivery time, which weren’t coherent to the distance i.e. a store 2.3 kms away showed 180 minutes as expected delivery time while that 11.6 kms away showed 65 minutes).</em></p><p><strong>Pain Point 4: </strong>Only 1 store tagged as <em>‘Top store’</em> can create a bias in the mind of the users and can think the store is being promoted.</p><p><strong>Pain Point 5: </strong>If the user wants to get their order delivered in the shortest time they would have to check the expected delivery time for each of the stores individually.</p><blockquote><strong><em>Recommendation 4: </em></strong>Ratings for each store should also be reflected, as 1 store being marked as ‘top rated’ and ‘wide variety’ might lead to bias in decision making. [Could have and missing]</blockquote><blockquote><strong><em>Recommendation 5: </em></strong>Filter and sort options should be given for the stores, given that the user might want to prioritize distance, rating or expected delivery time for stores. [Must have and missing]</blockquote><blockquote><strong><em>Additional Feature 1: </em></strong>Just below the search bar an option to add ‘Previously bought items’ could be suggested. This would help increasing the cart size/AOV. [Could have and missing]</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/553/1*vqgotxLsoNTx4Hr_Ip5nRQ.png" /><figcaption><strong>Fig. 7. Search page for items or stores</strong></figcaption></figure><p>The recommended search page looks something like Fig. 7. Once the user finalizes the store to order from, they can proceed to add items.</p><p>As soon as the user types in the product as shown in screen 3 in Fig. 3., list of stores with available matching items are displayed. As soon as the user adds an item <em>(in this case yogurt)</em> from any of the store they are automatically transitioned to a page with popular items in the store. And a small arrow as shown in screen 2 in Fig. 8 redirects to search within store page.</p><p><strong>Pain Point 6: </strong>The arrow might go unnoticed by the users and adds an additional step in user journey. The user might start a new search for another item and get a pop-up message for deleting earlier cart item as it is from another store in case where the newly searched item is unavailable in the same store.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*l80q93-cvqxYzckx-fWkaA.png" /><figcaption><strong>Fig. 8. Individual items search flow</strong></figcaption></figure><blockquote><strong><em>Recommendation 6: </em></strong>Instead of user clicking on the arrow labeled as ‘D’ in Fig. 8, automatic transition to screen 4 (i.e. search within store) should be done when the user click on add item in screen 1 (Fig. 8).</blockquote><p>The recommended flow would land onto the modified screen 4 in Fig. 8. as shown below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/613/1*SaTB9nzLkwvWv4UcsM95Ig.png" /><figcaption><strong>Fig. 9. Recommended individual items search flow</strong></figcaption></figure><p>Once the item has been added, the user proceeds for checkout by clicking on ‘<em>View Cart’ </em>button.</p><p>The purple banner at the bottom just above the <em>‘View Cart’</em> button suggests additional items to be added for free delivery. The positioning and the colour contrast makes it more noticeable and users can add to cart any item at this point to avoid coming back to this page later. <em>[Should have and done right]</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/922/1*kzN1AF33r1UVYfPharNRaA.png" /><figcaption><strong>Fig. 10. Checkout and payment screen</strong></figcaption></figure><p>Once the user is directed to the cart, at the top the savings are displayed in contrasting colour which instantly instills the pleasure of savings over guilt of spending. <em>[Could have and done right]</em></p><p><strong>Pain Point 7:</strong> <em>‘Add more’ </em>option can go unnoticed at the top right as it is mere text.</p><blockquote><strong><em>Recommendation 7: </em></strong>‘Add more’ option should be added after the items in the cart and made more noticeable as a button.</blockquote><p>The recommended screen is shown below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/340/1*__ukiK6dGm7b6tJk6ho_GQ.png" /><figcaption><strong>Fig. 11. Recommended ‘Add items’ button</strong></figcaption></figure><p>When the user click on <em>‘Pay’ </em>button, they are directed to page where payment gateways are listed with default selection as last used option. <em>[could have and done right].</em></p><p>However, as observed in screen 3 of Fig. 10, the JioOnePay interface seems a bit distorted and might lead to reduced user experience.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KxaqCgNVpVy2Xx4KJ-Wfdg.png" /><figcaption><strong>Fig. 12. Order tracking and cancellation</strong></figcaption></figure><p>Once the order is placed, user is directed to the tracking page.<em> ‘Dunzo Tales’ </em>is a really cool feature to keep the users engaged by conveying customer stories and hacks while they wait for their order. Making the wait enjoyable. <em>[Could have and done right]</em></p><p>Towards the very bottom of the page is the option to cancel accompanied with an empathetic message to ensure that the user doesn’t feel guilty on cancelling their order.</p><blockquote><strong><em>Additional Feature 2:</em></strong> Option to schedule order on the checkout page when everything to be bought is added to the cart. This would help user to order items like milk, egg, etc. that could be delivered early morning as they would be busy leaving for office and might not have time to order. [Could have and missing]</blockquote><blockquote><strong><em>Additional Feature 3:</em> </strong>Save repeat orders for user to order every month and display at the bottom of the home page so that the user can directly add those items to their cart instead of searching each item again. This will add a touch of personalization to the app. [Could have and missing]</blockquote><blockquote><strong><em>Additional Feature 4: </em></strong>The search feature should be improved to suggest items/stores closest to the text entered by the user. Ex. If the user typed “Aple” instead of “Apple”, the search result should not display error (as shown currently) but closest result. [Must have and missing]</blockquote><p><strong>Feature table</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/941/1*wHx-QJb3WnaBD39GL7eFwQ.png" /><figcaption><strong>Fig. 13. Feature comparison with close competitors</strong></figcaption></figure><p><strong>Metrics tracked to validate the success of above recommendations post implementation</strong></p><p>The following metrics if tracked could help us understand whether the changes made were a success or failure -</p><ol><li>Conversion rate</li><li>Repeat customer rate</li><li>Time to place an order</li><li>Average cart size</li><li>Search success rate</li><li>Daily/Monthly active users (DAU/MAU)</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ebd88dfc5777" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>