<?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 Debashis Nayak on Medium]]></title>
        <description><![CDATA[Stories by Debashis Nayak on Medium]]></description>
        <link>https://medium.com/@debashisnayak?source=rss-d5ee2eb07331------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*sX341HAU1rC0kyb8Trnglw.png</url>
            <title>Stories by Debashis Nayak on Medium</title>
            <link>https://medium.com/@debashisnayak?source=rss-d5ee2eb07331------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 20:05:01 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@debashisnayak/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[Caveman Cuts AI Output Tokens by 65%. It Saved Me Almost Nothing. · Products Decoded]]></title>
            <link>https://debashisnayak.medium.com/caveman-cuts-ai-output-tokens-by-65-it-saved-me-almost-nothing-products-decoded-0e630597ba8d?source=rss-d5ee2eb07331------2</link>
            <guid isPermaLink="false">https://medium.com/p/0e630597ba8d</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Debashis Nayak]]></dc:creator>
            <pubDate>Fri, 28 Aug 2026 00:00:52 GMT</pubDate>
            <atom:updated>2026-08-28T10:45:07.570Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*g-5qDyVfwhD4dEDp.jpg" /></figure><p>A tool that shrinks your agent’s replies by 65% sounds like an obvious install. Then you look at which side of the bill it actually touches. Input tokens, output tokens, and why the same tool can be a win for one setup and a loss for another.</p><p>I run most of my life through an AI setup now. Job search, three side products, family health tracking, a memory system that remembers what I told it six weeks ago. All of it costs tokens, and tokens cost money, and money is a thing I watch closely these days.</p><p>So when I found Caveman, I actually got a little excited. The pitch is great: make your AI coding agent talk like a caveman, same answers, 65% fewer output tokens. One command to install, works across Claude Code, Cursor, Gemini, thirty-odd agents. No backend, no telemetry, nothing phoning home. I installed it the same evening.</p><p>Then I did the thing I do for a living, which is read past the headline number. The honest answer is that Caveman is a good tool that will do very little for me. Not because it is broken. Because of how I use my AI, which turns out to be the whole game.</p><h3>What it actually is</h3><p>Caveman is not a proxy or a compression engine sitting between you and the model. It is a skill. A prompt. It tells your agent to drop the filler, write in fragments, skip the “Great question! Let me help you with that” throat-clearing, and keep code, commands, and error messages exactly as they are.</p><p>That is it. And that restraint is the smart part. It does not touch what the model can do. It only touches how much it says.</p><p>The benchmarks back the claim. Across ten coding tasks the average output shrank 65%, with a range from 22% on the low end to 87% on the high end. A React re-render explanation went from an essay to four lines. That is a real result, and if you have ever watched an agent write six paragraphs to tell you to add a key prop, you know exactly why people want this.</p><h3>Where the number gets slippery</h3><p>Here is what made me put the calculator down and actually think.</p><p>A model turn has two sides. There is what you send in, which is your files, your context, the diff, the previous conversation. And there is what the model sends back. Input tokens and output tokens. On most pricing, input is cheaper per token, but there is usually a lot more of it.</p><p>Caveman only shrinks the output side. It says so itself, and I respect that the project spells this out in its own honest-numbers notes instead of hiding it. The input tokens do not move. The reasoning tokens do not move. And the skill adds roughly 1,000 to 1,500 input tokens every single turn, because the instructions have to ride along each time.</p><p>So the 65% is a real cut on one slice of the bill. The question is how big that slice is for you.</p><p>For a lot of coding work, the input slice is the big one. You are feeding the model a repo, a stack trace, a long back and forth. The reply is often the small part. Cut 65% off the small part, add a fixed cost to the big part, and the whole-session saving comes out modest. On workloads that are already terse, the project admits it can even go slightly negative.</p><h3>Why my setup is the worst case for it</h3><p>This part is specific to me, and probably to anyone running a personal AI OS instead of one-off coding sessions.</p><p>My system injects context on every prompt. Before I even ask a question, a hook pulls the five most relevant memories out of a vector database and loads them in. Then my instruction file loads. Then my memory index loads. Every turn carries this weight, and it is all on the input side, which Caveman does not touch at all.</p><p>The output, meanwhile, is usually the smallest thing in the exchange. I ask a sharp question, I want a sharp answer, I have already trained the thing to be blunt with me. Caveman would be squeezing juice from the one part of my setup that is already dry, while quietly adding its 1,500-token tax to the part that is actually heavy.</p><p>For me the math does not just underwhelm. It runs the wrong way.</p><h3>The one feature I am still tempted by</h3><p>There is a piece called caveman-compress that rewrites your saved memory or context files into a denser form, and claims around 46% smaller input on future sessions. That one is interesting to me, because it hits input, which is my actual cost center.</p><p>But I am holding off, and here is the caution if your memory is anything like mine. My files are not just prose. They are wired into a semantic search that runs on embeddings, they carry structured links between notes, and they are full of exact figures I cannot afford to have rounded or dropped. Embedding models also read natural sentences better than caveman fragments, so compressing the text can quietly make retrieval worse even when the file looks fine to a human. Compress for readability and you can lose fidelity you did not know you were spending. If I ever run it, it goes on one throwaway copy first, with a careful diff after.</p><h3>Who should actually install this</h3><p>I do not want this to read as a takedown, because it is not one. Caveman is well built, honest about its own limits, and solves a real annoyance. Let me just be clear about the fit.</p><p>Install it if you do a lot of interactive coding, you are on output-priced or per-message limits, and you are tired of scrolling through padding to find the one command you needed. The speed and readability alone might be worth it, forget the cost. That is a genuine everyday win.</p><p>Skip it, or at least run the numbers first, if your setup is input-heavy. Long context windows, RAG pipelines, big system prompts, a memory layer that reloads every turn. That is where the fixed per-turn tax eats the output savings, and where you will stare at your bill next month wondering why it looks the same.</p><p>The tool is not the variable. Your token profile is. Caveman just made me go and actually look at mine, which is the most useful thing a tool that saved me nothing has done in a while.</p><p><em>Originally published at </em><a href="https://productsdecoded.com/blog/caveman-token-savings-input-vs-output"><em>https://productsdecoded.com</em></a><em> on August 28, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0e630597ba8d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Agents Are Live in Production. They’re Also Dying in Production.]]></title>
            <link>https://medium.com/design-bootcamp/ai-agents-are-live-in-production-theyre-also-dying-in-production-9f11b4ef6e62?source=rss-d5ee2eb07331------2</link>
            <guid isPermaLink="false">https://medium.com/p/9f11b4ef6e62</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Debashis Nayak]]></dc:creator>
            <pubDate>Mon, 27 Jul 2026 07:37:59 GMT</pubDate>
            <atom:updated>2026-07-27T07:37:59.185Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OBzsw5qDvZJsUG_9trxUAg.png" /></figure><p>On July 16th I stood on the Grand Road in Puri with my wife, holding a standee with a QR code on it. Rath Yatra crowd, lakhs of people, zero marketing plan. By evening my little AI astrology product, Astrika, had 87 paid users, roughly ₹8,000 in a single day. Most paid ₹49 for one question. A few bought the ₹249 full reading within hours of scanning, and a handful went deeper. That same night, the product started failing. One paying user at a time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JF2hrkk1LOzyWN_PHDg4hw.png" /></figure><p>I want to walk through what actually broke, because I think it’s the most honest picture of AI in production I can give you. Consider it a post-mortem, written by the person who caused the incident and then fixed it at 1 am.</p><h4>The bug that only shows up with real users</h4><p>Astrika.in answers Jyotish questions from a real birth chart. The engine ran on an AI model. During the traffic spike I hot-swapped the model to a newer version, mid-incident, without re-tuning the prompt. Seemed harmless. Bigger model, better answers, right?</p><p>Three things went wrong at once. The new model narrates more, so outputs that used to parse cleanly stopped parsing. It rejects parameters the old model accepted, so a single leftover `temperature: 0.3` in one service started returning 400 errors and silently killed every paid question on that path. And it asserts slightly different astrology, so the verification layer started flagging readings that were actually fine.</p><p>None of this appeared in testing. All of it appeared with paying users, one at a time, on the highest-traffic day the product ever had.</p><p>This is the production gap. The demo works and the pilot works, and then real users arrive with real money and the failure modes you never imagined start queuing up politely.</p><h4>Why agents die after launch</h4><p>I spent two years shipping AI at MoneyLion, including embedding live insurance offers inside Microsoft Copilot, and now I run my own AI products solo. Across both worlds, enterprise and one-person SaaS, the deaths look basically the same. Five patterns keep repeating.</p><p><strong>The model changes under you.</strong> Your LLM is a dependency that ships breaking changes without a changelog you actually read. New versions reject old parameters. Old versions get deprecated on someone else’s schedule. If your product treats the model as a constant, you have built on sand. My temperature bug was exactly this, a parameter that was fine for months until the model underneath moved.</p><p><strong>Real inputs are nothing like your test inputs.</strong> Your test prompts are written by someone who knows the product. Real users paste half a birth certificate, type in three languages, ask two questions in one, and expect one clean answer. The gap between your eval set and what actually arrives is where most agents quietly rot.</p><p><strong>Failures are silent.</strong> When a database call fails, you get an exception. When an LLM fails, you often get a confident, fluent, wrong answer. Nothing throws. The user just gets a bad reading, or a bad refund decision, or a bad summary, and you find out when they email you. Or worse, when they don’t.</p><p><strong>Nobody owns the failover.</strong> Ask an AI team what happens when their primary model returns 403 at 9 pm on a festival weekend. Most have no answer beyond “retry.” Retrying a dead model is just failing slower.</p><p><strong>Costs and latency compound.</strong> The demo answers one question. Production answers ten thousand, and suddenly the chain of calls you glued together in a hackathon costs more than the revenue it generates.</p><blockquote>Maybe there’s a sixth pattern about org politics, but these five are the ones I have personally bled on.</blockquote><h4>What shipping actually took</h4><p>Here’s what went into Astrika after that night. None of it is glamorous, which is kind of the point.</p><p>A failover chain instead of a single model. The engine now tries Model 1, then Model 2 (smaller), then another smaller model and then a backup model. It never falls back to the cheapest tier, because a paying user asking about their marriage deserves better than the budget model shrugging at them. A 403 or 404 triggers failover. A 400 does not, because a 400 means my request is malformed and switching models would just hide my own bug.</p><p>Structured outputs with a verifier. Every reading passes through seven deterministic checks before a user sees it. The chart math comes from a local <strong>Swiss Ephemeris engine</strong>, not from the LLM, so the model interprets but never calculates. LLMs are confident liars about arithmetic and planetary positions equally.</p><p>A smoke test I can run in one command. Before any deploy, `<strong>npm run smoke</strong>` fires real questions at the real chain and checks real outputs. Boring. Saved me at least three times since.</p><p><strong>An apology playbook. The users hit by the incident got an email and 2 free credits, automatically. Trust in an AI product is mostly recovered in the failure path, since the happy path is table stakes now.</strong></p><p>Status handling that doesn’t leak. My status endpoint used to expose which provider was down. Users don’t need to know my vendor drama, and competitors definitely don’t.</p><h4>The part that’s actually a PM job</h4><p>If you’re a PM moving into AI, this is the shift no course prepares you for. Your job is no longer “define the feature and ship it.” Your job is to manage a probabilistic dependency that changes monthly, and to decide, in advance, what happens when it misbehaves.</p><p>Concretely, that means writing evals like you used to write acceptance criteria. It means an error budget for prompts, because prompts regress exactly like code. It means treating a model swap as a full release with regression testing, never as a config change. I learned that last one the expensive way, in public, during a festival.</p><p>At MoneyLion the same discipline applied at a very different scale. When we put insurance offers inside Microsoft Copilot, the engineering was honestly the smaller half of the work. The larger half was proving the thing would behave predictably enough that a partner would attach their brand to it. Enterprise AI dies in procurement and review for the same reason my solo product died on Rath Yatra night: nobody trusts a system that fails in surprising ways.</p><h4>Demos are marketing, production is the product</h4><p>The AI feed rewards demos. Screen recordings of agents booking flights, refactoring repos, closing tickets. I enjoy them too. But a demo only has to work once, and production has to work at 1 am, during a traffic spike, on a model version that didn’t exist when you wrote the prompt.</p><p><strong>87 paid users in a day told me the product had a market. The incident that same night told me what the actual product is</strong>: the failover chain, the verifier, the smoke test, the apology email. The model is maybe 30 percent of it. I’m not fully sure about that number, but I’m sure it’s a minority share.</p><p>I still don’t have persistent logs on Astrika, by the way. It’s on the list. There’s always a list.</p><p><strong>I write weekly about shipping AI products that survive real users, from a $9.6M ARR vertical at MoneyLion to solo products built from my desk. Follow for the next one.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9f11b4ef6e62" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/ai-agents-are-live-in-production-theyre-also-dying-in-production-9f11b4ef6e62">AI Agents Are Live in Production. They’re Also Dying in Production.</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LinkedIn Is Feeding AI Content to AI Readers. Here’s What That Means for You]]></title>
            <link>https://debashisnayak.medium.com/linkedin-is-feeding-ai-content-to-ai-readers-heres-what-that-means-for-you-032efd204ae7?source=rss-d5ee2eb07331------2</link>
            <guid isPermaLink="false">https://medium.com/p/032efd204ae7</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Debashis Nayak]]></dc:creator>
            <pubDate>Sun, 07 Jun 2026 13:33:45 GMT</pubDate>
            <atom:updated>2026-06-07T13:33:45.650Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XGeuY4dEt30WWtJ00DCdRQ.jpeg" /></figure><p><strong>I scrolled LinkedIn for 15 minutes this morning. Here’s what I noticed:</strong></p><p>The posts were well-structured. The grammar was clean. The takes were formatted exactly the way LinkedIn takes are supposed to be formatted: hook, three observations, takeaway, question at the end.</p><p><strong>Almost none of it felt like it came from a person.</strong></p><p>The comments were supportive and specific-sounding. “<strong>This is such a great point, I’ve been thinking the same thing</strong>.” “<strong>Really valuable perspective, especially point 2.</strong>” They seemed engaged. Most of them were probably not written by anyone either.</p><p>We’ve built a content system where AI generates posts, AI tools comment on those posts, and the whole thing circulates through an algorithm that doesn’t know the difference between a genuine observation and a pattern-matched output that looks like one.</p><p><strong>This is a problem, but not for the reasons most people say.</strong></p><h4><strong>The obvious critique misses the point</strong></h4><p>The usual complaint is about authenticity: AI content is fake, it devalues the platform, it’s dishonest.</p><p>That framing doesn’t hold. Written content has always been shaped by editors, ghostwriters, templates, and tools. The line between “<strong>your voice</strong>” and “<strong>influenced by external patterns</strong>” has always been blurry.</p><p><strong>The real problem is different:</strong> when AI reads AI, the feedback loop produces content that looks more and more like content, and less and less like the thing that content is supposed to be about.</p><h4><strong>What the loop actually looks like</strong></h4><p>Someone generates a post about AI productivity tips. Well-structured, grammatically clean, formatted for the algorithm. It gets moderate reach because it fits the pattern the algorithm recognizes.</p><p>An AI comment tool reads the post and generates a comment that mirrors the energy: “<strong>Great insight! This really resonated</strong>.” That comment gets a like. Maybe a reply, also generated.</p><p>The original creator sees engagement. They generate another post in the same vein. The tool monitoring their account adds this post to its internal data for “<strong>what works</strong>.”</p><blockquote>Over time, the posts get more optimized for the algorithm pattern and less connected to any actual observation or experience.</blockquote><h4><strong>The signal problem for real creators</strong></h4><p>If you are trying to build a genuine audience, one that reads, thinks, and comes back because they trust your perspective, you are now competing in a feed where the algorithm cannot distinguish you from the AI posts.</p><p><strong>At least not yet.</strong></p><p>The thing the algorithm cannot measure, and that readers can still feel - is specificity. An AI-generated post about “<strong>5 lessons from my PM career</strong>” can be well-written and completely devoid of anything that happened to an actual person. A post that says “<strong>A hiring manager asked me to improve their card product’s spend categorization feature on the spot</strong> -<strong> here’s what I said</strong>” cannot be replicated. There is a specific company, a specific feature, a specific moment that only exists in one person’s memory.</p><p>Specificity is the signal that separates human from pattern-matched output.</p><h4><strong>What this means if you’re posting on LinkedIn</strong></h4><p>The bar for generic content is now zero, anyone can produce it in seconds. The bar for specific, experience-based content is as high as it’s always been: <strong>you have to have had the experience.</strong></p><p><strong>That’s where to put effort. Not in optimizing hook structure.</strong> Not in posting time. Not in hashtag research. In doing things that generate experiences worth writing about, and then writing about them with enough specificity that the content is uniquely yours.</p><blockquote>The AI content loop will keep producing volume. The readers who matter will keep filtering for the real thing. They always have, it’s just harder to find now.</blockquote><p>Follow me on Medium for more on content, AI, and building in public. I write 3 - 4 times a month.</p><p>#AI #LinkedIn #ContentCreation #ArtificialIntelligence #BuildingInPublic #ContentMarketing #ProductManagement #Technology</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=032efd204ae7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Everyone Is Vibe Coding. Almost No One Is Making Money From It]]></title>
            <link>https://blog.startupstash.com/everyone-is-vibe-coding-almost-no-one-is-making-money-from-it-4f6d6aca7b53?source=rss-d5ee2eb07331------2</link>
            <guid isPermaLink="false">https://medium.com/p/4f6d6aca7b53</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Debashis Nayak]]></dc:creator>
            <pubDate>Mon, 18 May 2026 19:31:00 GMT</pubDate>
            <atom:updated>2026-05-18T19:31:00.603Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FpsQCFpyUuLGRawkamvehQ.jpeg" /></figure><blockquote><strong>I’ve built more than 100 products across different businesses over the last 8 years.</strong></blockquote><blockquote><strong>The hardest part was never building. It was always the first paying customer.</strong></blockquote><blockquote><strong>That hasn’t changed. What’s changed is that everyone is building now -and almost nobody is making money.</strong></blockquote><p>In the last 6 months, I’ve watched a wave of vibe-coded products hit the internet. Founders post launch screenshots on LinkedIn. Products go live on Product Hunt. GitHub repos fill up with side projects. The building has never been faster or cheaper.</p><p>The monetization rate has not moved.</p><h4><strong>The validation step gets skipped entirely</strong></h4><p>Most vibe coders go straight from “<strong>I have an idea</strong>” to “<strong>I built the product</strong>.” The idea-to-ship cycle has compressed from months to days. That’s real, and it’s remarkable. But the thinking that used to happen during a longer build cycle - talking to users, understanding if someone will actually pay - doesn’t happen automatically just because the product ships faster.</p><p><strong>Speed doesn’t replace validation. It just makes it easier to skip.</strong></p><p>I’ve done this myself. At <strong>CoolHax Labs</strong>, I shipped products fast. Some of them were genuinely useful. A few found users. Almost none of them built a revenue line without deliberate, uncomfortable distribution work that had nothing to do with the product itself.</p><h4><strong>The building is the easy part now</strong></h4><p>This is the uncomfortable shift: the thing that used to create the moat - technical execution - is no longer the hard part. Claude can write the backend. Cursor can wire up the frontend. Vercel deploys it in 4 minutes.</p><p>What nobody has automated is the conversation where a stranger tells you why they will or won’t pay for the thing.</p><p>That conversation is still on you.</p><h4><strong>What separates the ones who monetize</strong></h4><p>I’ve looked at the vibe-coded products that actually generate revenue, and they share one thing: the builder talked to the customer before the build, during, and after launch. Not to validate the idea in a feel-good way. To find out what the customer was already paying for, and whether this product could displace that spend.</p><p>The ones that fail almost always made an assumption about pricing at the start and never tested it until they had a full product waiting for someone to buy.</p><h4><strong>The specific mistake: building for usage, not payment</strong></h4><p>There’s a difference between “<strong>people are using this</strong>” and “<strong>people are paying for this.</strong>”</p><p>Free usage feels like validation. It is not.</p><p>If someone uses a product for free and stops when you add pricing, that’s not a product-market fit problem - it’s a value problem. They found it useful but not useful enough to pay for.</p><p>Vibe coding makes it easy to get to free usage fast. It doesn’t help you understand the gap between “<strong>useful</strong>” and “<strong>worth paying for</strong>.”</p><h4><strong>What to do instead</strong></h4><p>Before the next build, answer this: <strong>who has this problem badly enough that they are currently paying for a worse solution?</strong></p><p>If you can name three people - not personas, not “<strong>small business owners in their 30s</strong>,” but actual people - you have something to build toward. If you can’t, build anyway, but give yourself a hard deadline to find those three people before you write another line of code.</p><p><strong>The tool is not the product. The paying customer is the product.</strong></p><p>Follow me on Medium for more on building, PM, and what actually works with AI. I write 3 - 4 times a month.</p><p>#ProductManagement #AI #VibeCoding #Startups #BuildingInPublic #Founder #SaaS #AITools</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4f6d6aca7b53" width="1" height="1" alt=""><hr><p><a href="https://blog.startupstash.com/everyone-is-vibe-coding-almost-no-one-is-making-money-from-it-4f6d6aca7b53">Everyone Is Vibe Coding. Almost No One Is Making Money From It</a> was originally published in <a href="https://blog.startupstash.com/">Startup Stash</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Almost Paid for wisprflow. Then I Built My Own With Gemma 3 1B.]]></title>
            <link>https://debashisnayak.medium.com/i-almost-paid-for-wisprflow-then-i-built-my-own-with-gemma-3-1b-1cd4558dced6?source=rss-d5ee2eb07331------2</link>
            <guid isPermaLink="false">https://medium.com/p/1cd4558dced6</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Debashis Nayak]]></dc:creator>
            <pubDate>Sun, 17 May 2026 09:12:34 GMT</pubDate>
            <atom:updated>2026-05-17T09:12:34.184Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pKOmoBqF-C-exVDm4cmiSw.png" /></figure><p>I was on the <a href="https://medium.com/u/57a7a54a594b">Wispr Flow</a> checkout page.</p><p>Annual subscription. For a dictation app.</p><p>I closed the tab.</p><p>Not because wisprflow is bad. It’s genuinely well-built. But I’d spent 15 minutes on the checkout page calculating whether I’d use it enough to justify the cost. The moment I start that calculation, I already have my answer.</p><p>So I started looking for alternatives. Two days later, I had Voice Link running on my MacBook - local, free, and producing the same(almost similar i would say) output.</p><h4><strong>What I was trying to solve</strong></h4><p>Wisprflow does one thing well: you speak, it transcribes, the text appears wherever your cursor is. System-wide. No switching windows, no copying text from a separate app. Just talk.</p><p>I wanted exactly that. I did not want to pay a subscription for it indefinitely.</p><h4>The exploration</h4><p>The obvious starting point was running Whisper locally. OpenAI’s Whisper is open source, several Mac apps are built on top of it, some free. I tried three of them. Most had friction I didn’t want to deal with: clunky setup, menu bar apps that felt half-finished, or transcription quality that degraded with background noise.</p><p>Then I looked at what was available for local language models on Apple Silicon.</p><p><strong>Gemma 3 1B</strong> caught my attention. Google’s Gemma 3 at 1 billion parameters is small enough to run entirely on CPU and Neural Engine - no GPU required, no cloud connection. On an M1 MacBook Pro it runs fast. Fast enough to be useful in real time, not just as a novelty.</p><blockquote>I hadn’t seen anyone use it for transcription cleanup. I tried it.</blockquote><h4>Building Voice Link</h4><p>The stack I ended up with:</p><p>Whisper small (local) handles the speech-to-text conversion. It catches the words accurately but produces raw output - no punctuation, occasional run-on sentences, spoken filler words left in.</p><p><strong>Gemma 3 1B takes that raw output and cleans it</strong>. Adds punctuation, removes “uh” and “um”, fixes sentences that make sense spoken but read awkwardly written.</p><p>It’s a short prompt: <strong><em>clean up this transcription while preserving the speaker’s meaning. That’s it.</em></strong></p><p>A lightweight Mac menu bar app ties the two together and pipes the final text to wherever the cursor is.</p><p><strong>Total build time:</strong> parts of two days. The actual coding was half day. The rest was tuning the Gemma prompt until the cleanup felt natural rather than over-edited.</p><h4>What surprised me</h4><p>The quality gap is smaller than I expected.</p><p>Wisprflow has clearly been iterated on. It handles accents, domain-specific vocabulary, and background noise better than my setup. For someone who dictates 4 - 6 hours a day, that polish matters. I’m not dismissing it.</p><p>But for occasional dictation, messages, first drafts, notes - Voice Link is close enough that the gap is not worth a subscription.</p><p>What genuinely surprised me was Gemma 3 1B doing the cleanup job well. I expected it to over-correct or introduce errors. Instead it improved the transcription without changing meaning. A 1 billion parameter model, running entirely locally, making spoken text readable.</p><p><strong><em>That was not what I expected from a model this small.</em></strong></p><h4>The shift that made this possible</h4><p>Two years ago, running a useful local LLM on a MacBook meant a 7B model minimum, careful quantization, and slow inference. 1B models were not worth considering for anything generative.</p><p>Gemma 3 1B is different. It’s fast enough, accurate enough, and small enough to run as a background process without affecting the rest of the machine.</p><p><strong>The implication:</strong> tools that were subscription-dependent because they required cloud inference are now buildable locally. Not for every tool. Not for every use case. But for utility apps where the core function is a single, repeatable task, transcription cleanup, text classification, quick summarization - the math has changed.</p><p>The model is free. The compute is on your machine. The subscription is a choice, not a requirement.</p><h4>What I’m still paying for</h4><p>I’m not suggesting everyone stop paying for software. There are tools I pay for where the value is clear: network effects, constant model updates, polish I’d need weeks to rebuild.</p><p>But for tools where the core function is a commodity model doing a commodity task - I’m building before I subscribe now. Voice Link took two days. The subscription it replaced costs more per year than I’d spend in time rebuilding it.</p><p>That calculus is new. A year ago it wasn’t true. Now it is.</p><p>Follow me on Medium for more on building with AI and what actually works on Apple Silicon. I write 3 - 4 times a month.</p><p>#AI #MachineLearning #BuildingInPublic #Gemma #AppleSilicon #ProductManagement #AITools #OpenSource</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1cd4558dced6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[If You Don’t Use Your Own Product, You’ve Already Answered the Question]]></title>
            <link>https://medium.com/design-bootcamp/if-you-dont-use-your-own-product-you-ve-already-answered-the-question-5a4fefeb96df?source=rss-d5ee2eb07331------2</link>
            <guid isPermaLink="false">https://medium.com/p/5a4fefeb96df</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Debashis Nayak]]></dc:creator>
            <pubDate>Fri, 15 May 2026 22:21:24 GMT</pubDate>
            <atom:updated>2026-05-15T22:21:24.150Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r_hglea9UUPFhse10DYu4Q.jpeg" /></figure><p>A founder I know built a B2B SaaS tool for sales teams. 12 months in, decent revenue, growing slowly.</p><p><strong>I asked him:</strong> do you use it to manage your own sales pipeline?</p><p>“<strong>It’s not really built for someone like me</strong>,” he said.</p><p><strong>That sentence does a lot of work.</strong></p><p>If the person who spent a year building the product, who knows every feature, who wrote the copy explaining why it’s useful, if that person won’t use it, the question of why customers churn writes itself.</p><h4><strong>The uncomfortable reason founders stop using their own products</strong></h4><p>It’s not usually because the product is bad. It’s because they know where all the bodies are buried.</p><p>Founders know which features are half-built. They know which edge cases break. They’ve seen the Slack thread where the bug got deprioritized for three sprints. So they route around those parts and eventually stop using the product to avoid the frustration.</p><p><strong>The problem:</strong> by doing this, they’ve cut themselves off from exactly the information they need.</p><p>Every time a real user hits one of those broken edges, they experience something the founder has quietly learned to avoid. The founder doesn’t feel the friction because they’ve automated around it. <strong>The user churns because they haven’t.</strong></p><h4><strong>What actually using your own product looks like</strong></h4><p>It’s not ceremonial. It’s not logging in once a month to check the dashboard loads.</p><blockquote>It means using the product under the same constraints your users face, same information, same flow, same onboarding. No admin shortcuts. No backdoors. No asking the engineer to manually fix your account.</blockquote><p>I use Jobtune for my own job applications. Not a demo account. My actual applications, with my actual resume, against real job descriptions.</p><p>Last week the AI roast told me to cut my insurance product experience because it “<strong>adds noise</strong>.” That experience is my strongest differentiator for fintech roles. The model didn’t know that.</p><p><strong><em>I fixed the prompt the next day.</em></strong></p><p>That fix came from using the product, not from a user complaint or a support ticket. From feeling the exact frustration a job-seeking PM would feel, because I am one.</p><h4><strong>The trust signal customers can feel</strong></h4><p>There’s a version of this that’s purely about product quality. But there’s also a trust signal that matters to customers.</p><p>When you say “<strong>I use this myself</strong>,” they can tell whether that’s true or a line. Real users ask specific questions: “<strong>I noticed that when you do X, it does Y, is that intentional?</strong>” Founders who actually use the product answer those immediately. Founders who don’t hedge.</p><p><strong>Customers notice the difference. It’s not subtle.</strong></p><h3><strong>The hard part</strong></h3><p>Using your own product is uncomfortable. You will find things that don’t work. You will get frustrated. You will want to ask your engineer to fix your specific account so you don’t have to deal with it.</p><p>Don’t. The friction you feel is data. It’s the same friction every user hits before they decide whether to come back.</p><p>The question isn’t whether your product is perfect. It’s whether you’re willing to feel what your users feel.</p><p><strong>If the answer is no, you already know what the retention problem is.</strong></p><p>Follow me on Medium for more on product thinking and building in public. I write 3 - 4 times a month.</p><p>#ProductManagement #SaaS #BuildingInPublic #Founder #Startups #ProductThinking #PM #Jobtune</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5a4fefeb96df" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/if-you-dont-use-your-own-product-you-ve-already-answered-the-question-5a4fefeb96df">If You Don’t Use Your Own Product, You’ve Already Answered the Question</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Fed 4 Years of Family Medical Reports Into Claude. AI Found What No Doctor Had Connected.]]></title>
            <link>https://debashisnayak.medium.com/i-fed-4-years-of-family-medical-reports-into-claude-ai-found-what-no-doctor-had-connected-2043a1169ca3?source=rss-d5ee2eb07331------2</link>
            <guid isPermaLink="false">https://medium.com/p/2043a1169ca3</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[mental-health]]></category>
            <category><![CDATA[family]]></category>
            <dc:creator><![CDATA[Debashis Nayak]]></dc:creator>
            <pubDate>Sat, 09 May 2026 12:23:35 GMT</pubDate>
            <atom:updated>2026-05-09T12:23:35.259Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FzUQXG7vA-oT1LAZ0lAIxg.png" /></figure><p>Last month I spent a Sunday afternoon feeding 4 years of family medical reports into Claude.</p><p>Blood tests, prescriptions, specialist notes, lab values across three family members. My mom’s rheumatology reports. My dad’s cardiac panels. My wife’s hormone and metabolic labs. I’d been meaning to organize all of it for months. What I didn’t expect was that organizing it would turn into something close to a diagnosis.</p><h4><strong>The Setup</strong></h4><p>My mom has been managing what her doctors describe as joint pain, fatigue, and general inflammation for years. She has three separate specialists:</p><ul><li>a rheumatologist</li><li>an endocrinologist</li><li>and a general physician.</li></ul><p>Each sees her every few months. Each runs their tests, adjusts their prescriptions, and sends her home with advice that sometimes contradicts the others.</p><p>On paper, she’s being looked after by three doctors. In practice, no single doctor has ever seen all her reports together.</p><p><strong>I built a simple tracker:</strong> a structured folder organized by person and date, every report scanned or photographed, fed into a Claude conversation with one prompt asking it to find patterns across conditions and time.</p><p><strong>It took Claude about 40 seconds.</strong></p><h4>What It Found</h4><p>My mom has confirmed rheumatoid arthritis, hypothyroidism, and mild anemia. Three separate diagnoses, managed by three separate doctors.</p><p>Claude pointed out something none of them had connected across 4 years: these three conditions compound each other in a documented feedback loop. RA causes anemia through a mechanism called anemia of chronic disease. Hypothyroidism worsens joint inflammation and fatigue, which makes RA symptoms harder to separate from thyroid symptoms. Treating each condition in isolation, without adjusting for the others, means the dosing is always slightly off.</p><p><strong>No doctor was wrong.</strong> The system just doesn’t give them the full picture. A rheumatologist sees one visit, one set of labs, one condition. An AI given 4 years of reports sees the trajectory.</p><p><strong>For my dad:</strong> his LDL has been tracked across 3 different labs over 2 years. The numbers looked acceptable in isolation at each visit. Plotted across time, there’s a clear upward trend that nobody flagged because no single comparison triggered an alarm. His current value is above what cardiac guidelines recommend for someone his age.</p><p><strong>For my wife:</strong> a hormone panel from 8 months ago had a marker that, combined with her current HOMA-IR reading, points toward early insulin resistance that wasn’t part of any conversation she’d had with her doctor. Not dramatic. But early enough to matter if caught now.</p><p>All three had patterns in their data that only became visible when someone looked at more than one snapshot at a time.</p><h4>What I Actually Did With It</h4><p>I’m not a doctor. I didn’t change anyone’s prescriptions.</p><p>I printed the AI’s summary, highlighted the specific cross-condition findings, and sent each family member to their respective doctor with pointed questions. “<strong>Can you look at these two reports together and tell me if this pattern is worth addressing?</strong>”</p><p><strong>Two out of three doctors agreed it was.</strong></p><p>One of them said: “<strong>We don’t usually have time to go back through years of reports in a consultation.</strong>” That sentence stayed with me longer than anything else from that afternoon.</p><h4><strong>The Real Problem</strong></h4><p>Healthcare is fragmented by design. Each specialist is optimized for their domain. The consultation is 15 minutes. The records are in different portals, sometimes different hospitals, sometimes printed and sitting in a folder at home that nobody’s touched in two years.</p><p><strong>The problem isn’t bad doctors.</strong> The problem is that the data required to see the full picture is scattered, and nobody’s job is to put it together before the appointment.</p><p>That used to be an unsolvable problem. It isn’t anymore.</p><p>Any capable model can read 50 pages of medical reports in seconds. It can cross-reference lab values across years. It can flag a trend that only becomes visible when you look at reports from three different labs side by side. It doesn’t replace the doctor’s judgment. It can prepare you to have a better conversation with one.</p><p>The doctor still makes the call. But you walk in with a question they can actually answer, instead of a vague description of how you’ve been feeling lately.</p><h4>How I Set It Up</h4><p>The tracker is simple. A folder structure organized by person, then by date. Each report goes in as a PDF or a scanned image. Every month or so, I run one prompt: “<strong>You have access to these reports. Find any patterns that cross conditions or time periods. Flag anything that has changed significantly in the last 6 months.</strong>”</p><blockquote><strong>That’s it.</strong></blockquote><p>No app. No code. Just organized information and a question that nobody else had thought to ask across the whole dataset at once.</p><h4>What’s Worth Trying</h4><p>If you have family members with chronic conditions managed by different doctors, or multiple conditions that have never been looked at together, or years of scattered lab reports sitting in a folder somewhere:</p><p>Collect everything you can find. Scan it. Feed it into a conversation with your AI tool of choice. Ask it to look for patterns across conditions and across time. Print whatever it finds that seems relevant. Take it to the doctor and ask for a response.</p><p>You’re not replacing the doctor. You’re doing the preparation work that the current system doesn’t have the time or structure to do before the appointment.</p><p>The AI isn’t the diagnosis. It’s the second set of eyes on data that nobody else was looking at in full.</p><p>That’s enough to change what happens in the room.</p><p>I write about AI, building in public, and what it actually looks like to use these tools in real life. If this was useful, follow me on Medium. I publish 3-4 times a month.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2043a1169ca3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Cloned My Voice for Free on a MacBook. ElevenLabs Charges $22/Month for the Same Thing.]]></title>
            <link>https://debashisnayak.medium.com/i-cloned-my-voice-for-free-on-a-macbook-elevenlabs-charges-22-month-for-the-same-thing-bd8d4caa637e?source=rss-d5ee2eb07331------2</link>
            <guid isPermaLink="false">https://medium.com/p/bd8d4caa637e</guid>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Debashis Nayak]]></dc:creator>
            <pubDate>Sun, 19 Apr 2026 18:06:00 GMT</pubDate>
            <atom:updated>2026-04-19T18:06:00.162Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>A step-by-step guide to local voice cloning with VoxCPM2 — no API keys, no subscriptions, no cloud.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jRbWDPghPvexl9Yt_M4JDA.jpeg" /></figure><p>Three months ago I was paying ElevenLabs $22 a month to generate voiceovers. Last week I cloned my own voice locally, for free, on a MacBook M1, using an open source tool called <strong>VoxCPM2</strong>. The audio quality is not identical. But it’s close enough that I used it to narrate a real reel - and I’m not going back.</p><p>This is exactly how I did it.</p><h3>Why I stopped paying ElevenLabs</h3><p>ElevenLabs is genuinely good. The cloned voices are clean, the interface is simple, and the API is well-documented. But <strong>$22/month</strong> for a side project that generates maybe 4 - 5 voiceovers a month is hard to justify - especially when you’re building a product on your own dime.</p><p>I started looking for alternatives. Most open source TTS tools give you robot voices. A few give you decent voices but require a GPU. VoxCPM2 runs on CPU, works on an M1, and does actual voice cloning - not just generic text-to-speech.</p><h3>What VoxCPM2 actually does</h3><p>VoxCPM2 is an open source voice cloning model. You give it a reference audio file of your voice, you give it text, and it generates speech in your voice. No API keys. No cloud. Everything runs locally.</p><p>It’s not magic. The quality depends heavily on: - Length and cleanliness of your reference audio - Whether you use a style prefix in your prompt, How you normalize the output after generation</p><p>I’ll cover all three.</p><h3>Step 1: Install VoxCPM2</h3><p>You need Python 3.11. I use conda environments to keep things clean.</p><pre>conda create -n voiceclone python=3.11<br>conda activate voiceclone<br>pip install voxcpm</pre><p>That’s it. The model weights download on first run. On M1 with a 100Mbps connection, expect about 5 - 8 minutes.</p><p><strong>CPU throttle (important on M1):</strong> Without this, VoxCPM will use all cores and your MacBook fans will sound like a jet.</p><pre>export OMP_NUM_THREADS=5<br>export MKL_NUM_THREADS=5<br>export OPENBLAS_NUM_THREADS=5<br>export PYTORCH_NUM_THREADS=5</pre><p>Set these before every generation command. With 5 threads on a 10-core M1, CPU usage sits at ~50–56%. Comfortable for background generation while you work.</p><h3>Step 2: Build your reference audio</h3><p>This is the most important step. The model learns your voice from the reference audio - garbage in, garbage out.</p><p><strong>What you need:</strong></p><ul><li>At least 60 seconds of your voice. 2–3 minutes is better.</li><li>Clean audio. No background music, no noise, no room echo.</li><li>Mono, 16kHz. VoxCPM2 will accept other formats but performance drops.</li></ul><p>I recorded five short clips on my MacBook’s built-in mic (not ideal - more on that later), then merged and resampled them:</p><pre>import soundfile as sf<br>import numpy as np<br>from scipy.signal import resample_poly<br>from math import gcd</pre><pre>files = [&#39;one.wav&#39;, &#39;two.wav&#39;, &#39;three.wav&#39;, &#39;four.wav&#39;, &#39;five.wav&#39;]<br>target_sr = 16000<br>chunks = []</pre><pre>for f in files:<br>    data, sr = sf.read(f)<br>    if data.ndim &gt; 1:<br>        data = data.mean(axis=1)  # stereo to mono<br>    g = gcd(target_sr, sr)<br>    resampled = resample_poly(data, target_sr // g, sr // g)<br>    chunks.append(resampled)</pre><pre>merged = np.concatenate(chunks)<br>sf.write(&#39;speaker.wav&#39;, merged, target_sr, subtype=&#39;PCM_16&#39;)</pre><p>My final speaker.wav ended up at 142 seconds. That&#39;s enough for the model to capture tone, pacing, and cadence.</p><h3>Step 3: Generate your first clone</h3><pre>conda run -n voiceclone bash -c &#39;<br>export OMP_NUM_THREADS=5 MKL_NUM_THREADS=5 OPENBLAS_NUM_THREADS=5 PYTORCH_NUM_THREADS=5<br>voxcpm clone \<br>  --text &quot;Your text here.&quot; \<br>  --reference-audio /path/to/speaker.wav \<br>  --output /path/to/output_raw.wav<br>&#39;</pre><p>The first time I ran this, the output sounded robotic. Flat. Like a voice actor reading a legal document.</p><p>The fix: <strong>use a style prefix.</strong></p><pre>--text &quot;(warm, conversational, natural pace, expressive) Your text here.&quot;</pre><p>That parenthetical at the start changes the generation significantly. It tells the model how to deliver the text, not just what to say. Without it, the default delivery is stiff. With it, the voice breathes.</p><p>Other formatting cues that work:</p><ul><li>Commas → short pause</li><li>... → longer pause, trailing thought</li><li>Em-dashes → emphasis before the next word</li><li>Short punchy sentences → force natural full-stop pauses</li></ul><h3>Step 4: Fix the volume (every time)</h3><p>Every raw output from VoxCPM2 comes out quiet - typically 4-7% peak amplitude. Barely audible without headphones. You have to normalize after every generation.</p><pre>import soundfile as sf<br>import numpy as np</pre><pre>data, sr = sf.read(&#39;output_raw.wav&#39;)<br>normalized = data * (0.95 / np.max(np.abs(data)))<br>sf.write(&#39;output_final.wav&#39;, normalized, sr, subtype=&#39;PCM_16&#39;)</pre><p>This brings the peak to 95% - loud enough to use without distortion. I ran this script after every single generation. Make it a habit.</p><h3>What I actually generated</h3><p>I used this setup to narrate an Instagram reel script for my “8 AI Agents” article. The script was 29.8 seconds of audio, exactly the length I needed. Here’s the script I fed it:</p><blockquote>“I typed one message to my AI at 11pm. ‘Audit my startup before I launch tomorrow.’ Then I went to sleep. By 7am… eight AI agents had filed 50 bugs I never caught. A race condition letting users steal credits. An IDOR flaw exposing private data. A legal consent checkbox… that did absolutely nothing.”</blockquote><p>The output wasn’t perfect. My built-in MacBook mic picks up room reverb that bleeds into the reference audio, and you can hear it in the clone. But it was usable, it was mine, and it cost nothing.</p><p><strong>Here is a sample generated by VoxCPM</strong></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fw.soundcloud.com%2Fplayer%2F%3Furl%3Dhttps%253A%252F%252Fapi.soundcloud.com%252Ftracks%252F2304510902%26show_artwork%3Dtrue&amp;display_name=SoundCloud&amp;url=https%3A%2F%2Fsoundcloud.com%2Fdebashis-nayak-985268487%2Fclone-sample%3Fsi%3Db24cee762fae481d8948a8532414b62f%26utm_source%3Dclipboard%26utm_medium%3Dtext%26utm_campaign%3Dsocial_sharing&amp;image=https%3A%2F%2Fsoundcloud.com%2Fimages%2Ffb_placeholder.png&amp;type=text%2Fhtml&amp;schema=soundcloud" width="800" height="166" frameborder="0" scrolling="no"><a href="https://medium.com/media/b0b57af8edb18eecc141aef45ca7dfbd/href">https://medium.com/media/b0b57af8edb18eecc141aef45ca7dfbd/href</a></iframe><h3>The honest limitation</h3><p>The reference audio quality is the ceiling.</p><p>I recorded on a MacBook built-in mic because my DJI Mic 1 wasn’t connecting (turned out to be a charge-only USB-C cable - cost me an afternoon). Clean close-mic audio with a proper lavalier or condenser mic would have produced noticeably better clones.</p><p>If you have a decent external mic, your results will be better than mine. If you’re recording in a quiet room on a laptop mic, you’ll get what I got - passable, not studio-grade.</p><p>The --prompt-text flag (Ultimate Cloning mode) takes this further. If you provide the exact transcript of what was spoken in the reference audio, the model can align phonemes more precisely. I haven&#39;t tested this fully yet, but the quality difference is reportedly significant.</p><h3>ElevenLabs vs VoxCPM2 — the honest comparison</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/952/1*YskB1Y5GQd7X7kWo3pS3Hw.png" /></figure><p>If you’re producing professional podcast intros or client-facing content and quality is non-negotiable - use ElevenLabs. If you’re a builder, experimenting, or just don’t want to pay $264/year to speak in your own voice - this setup works.</p><h3>The setup in under 5 minutes</h3><pre># Install<br>conda create -n voiceclone python=3.11 -y<br>conda activate voiceclone<br>pip install voxcpm soundfile numpy scipy</pre><pre># Record 2–3 minutes of your voice, then build reference:<br># (run the merge script above → speaker.wav)</pre><pre># Clone<br>conda run -n mlcourse bash -c &#39;<br>export OMP_NUM_THREADS=5 MKL_NUM_THREADS=5 OPENBLAS_NUM_THREADS=5 PYTORCH_NUM_THREADS=5<br>voxcpm clone \<br>  --text &quot;(warm, conversational, natural pace) Your text here.&quot; \<br>  --reference-audio speaker.wav \<br>  --output raw.wav<br>&#39;</pre><pre># Normalize<br>python3 -c &quot;<br>import soundfile as sf, numpy as np<br>d, sr = sf.read(&#39;raw.wav&#39;)<br>sf.write(&#39;final.wav&#39;, d * (0.95/np.max(np.abs(d))), sr, subtype=&#39;PCM_16&#39;)<br>&quot;</pre><p>Your voice. Your hardware. Zero dollars.</p><p><em>I write about building AI tools in public, PM career stories, and open source setups that replace expensive SaaS. Follow me on Medium if that’s useful to you.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bd8d4caa637e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Ran 8 AI Agents Overnight on My Own Startup. They Found 50 Bugs I Missed.]]></title>
            <link>https://medium.com/design-bootcamp/i-ran-8-ai-agents-overnight-on-my-own-startup-they-found-50-bugs-i-missed-2f941104a615?source=rss-d5ee2eb07331------2</link>
            <guid isPermaLink="false">https://medium.com/p/2f941104a615</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Debashis Nayak]]></dc:creator>
            <pubDate>Thu, 16 Apr 2026 09:48:21 GMT</pubDate>
            <atom:updated>2026-04-16T09:48:21.362Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S5JuR_LDe1xoAwnvQ35jSA.jpeg" /></figure><p><strong>What happens when you audit your own product with the same AI stack you used to build it.</strong></p><p>At 11pm on a Tuesday, I typed a single message to Claude: <em>“</em><strong><em>Run a full audit of Jobtune.io before I launch tomorrow.</em></strong><em>”</em></p><p>Then I went to sleep.</p><p>By 7am, eight AI agents had filed 50 issues across my codebase - security vulnerabilities, legal compliance gaps, broken copy, a race condition that let users steal credits from each other, and an IDOR flaw I’d introduced and never caught. I was the PM, the developer, and the QA team simultaneously.</p><p>Turns out I wasn’t good enough at all three.</p><h3>Building alone is cognitively expensive</h3><p>In March 2026, I stepped away from my role at MoneyLion and started building. Six weeks later, I had a live product - <strong>Jobtune.io</strong>, an AI resume roaster that tells job seekers exactly why their resume isn’t converting. I built it alone: product, design, backend (Node/Express/MongoDB), frontend (Next.js), payments (Razorpay), auth (Clerk).</p><p>I reviewed my own code. I tested my own flows. I had the blind spots of someone who is too close to the work and too tired to notice them.</p><p>The night before launch, instead of deploying and hoping, I decided to run the most thorough QA I could - not by staying up myself, but by setting agents on it while I slept.</p><h3>The audit stack: 8 specialized agents in parallel</h3><p>I used Claude (claude-sonnet-4.6) via Claude Code, running 8 background agents in parallel. Each had a specific, narrow scope:</p><ol><li><strong>Security agent</strong> - auth, IDOR, input sanitization, rate limits</li><li><strong>Billing integrity agent</strong> -<strong> </strong>Credit deduction logic, payment race conditions</li><li><strong>Legal/compliance agent</strong> -<strong> </strong>DPDP, GDPR patterns, consent flows</li><li><strong>UX copy agent</strong> -<strong> </strong>error messages, CTAs, empty states, pricing page copy</li><li><strong>QA functional agent</strong> - broken flows, dead links, missing fallbacks</li><li><strong>API consistency agent</strong> - route patterns, error response formats</li><li><strong>Frontend logic agent</strong> -<strong> </strong>state management, conditional renders, async handling</li><li><strong>Infrastructure agent</strong> - env vars, logging, secrets in source</li></ol><p>Each agent read the full codebase, filed structured findings with severity ratings (Critical / High / Medium / Low), and stopped. No fixes. Just findings.</p><h3>The 50 issues: what they tell you about solo founder blind spots</h3><p>By 7am I had a list. Here are the ones that would have hurt most:</p><p><strong>The race condition that let users steal credits</strong></p><p>My credit deduction logic ran <em>after</em> the OpenAI API call. The sequence was: <strong>hit endpoint → call OpenAI → deduct credits → save result</strong>. Anyone who knew the timing could fire parallel requests, receive multiple AI responses, and be charged for one. In a product where credits are the revenue unit, this was existential. One-line fix. I’d been shipping it for six weeks.</p><p><strong>The IDOR on interview audio</strong></p><p>One route — /api/interviews/:id/speak — fetched interview data by ID without checking ownership. Pass any valid interview ID and you could access someone else&#39;s audio. I&#39;d fixed this exact pattern on five other routes. I&#39;d missed it here.</p><p><strong>Sentry DSN hardcoded in source</strong></p><p>My error tracking key was in the source file, not in an environment variable. Visible to anyone who inspected the bundle. Not an immediate catastrophe — but embarrassing and wrong.</p><p><strong>The consent overlay that didn’t actually block anything</strong></p><p>My privacy consent overlay applied pointer-events: none to the Clerk signup form underneath it. The form looked blocked but was technically submittable. The warning text was also white on a white background — invisible. I had built a legal compliance checkbox that did nothing.</p><p><strong>13 broken copy issues</strong></p><p>Wrong pricing labels. A button running alert() in production. An interview score heading showing the same message whether you scored 90 or 30. These were the things users would notice in the first five minutes.</p><h3>The fix protocol: triage before touching anything</h3><p>The instinct after seeing 50 issues is to fix everything at once. That’s how you introduce new bugs.</p><p>I triaged against three questions: does it cost a user money or trust? Can it be exploited immediately? Does it break the first 10-minute experience? That gave me four buckets:</p><ul><li><strong>Fix before launch, no discussion:</strong> race condition, IDOR, hardcoded secret, consent overlay</li><li><strong>Fix within 24 hours:</strong> all 13 copy issues, dead links, broken routing</li><li><strong>Fix before 100 users:</strong> legal compliance gaps, credit refund-on-failure logic</li><li><strong>Fix when load becomes real:</strong> performance, database indexes, async job queues</li></ul><p>Six commits over two days. Launched on the seventh.</p><h3>The uncomfortable truth about AI QA</h3><p>Most early-stage startups don’t have a QA team. They have a founder who tests their own product the night before shipping. That founder is optimistic, sleep-deprived, and wrong about what they’re looking for.</p><p>Eight AI agents running for eight hours found more real bugs than I’d have caught in three weeks of manual testing — not because they’re smarter, but because they aren’t me. They don’t share my assumptions. They don’t skip testing the things I “know work.” They don’t get tired.</p><p>The race condition? I knew credit deduction happened “somewhere after the API call.” I’d never questioned the order. The billing agent had no assumptions. It read the code, compared it against billing integrity patterns, and flagged it immediately.</p><p>This is the actual leverage point. Not that AI writes better code than you. It’s that AI doesn’t inherit your blind spots.</p><h3>What I’d do differently</h3><p>Run the audit earlier. Not the night before launch — two weeks before. The race condition was a one-line fix. If I’d found it at launch I’d have been patching production under pressure with users watching.</p><p>The other thing: specialized agents dramatically outperform generalist prompts. One agent asked to “review security” gives you a surface scan. An agent scoped to “check every route for IDOR vulnerabilities using the ownership check pattern already in this codebase” finds real issues.</p><p>Write the agent brief like you’re onboarding a senior engineer who has never seen your code. The more specific the mandate, the better the output.</p><h3>The setup if you want to try this</h3><p>Minimum viable audit stack: one security agent, one billing/business logic agent, one UX copy agent. Three agents, one night. That’s the floor.</p><p>You built the product. You’ve been too close to it for months. Let something that’s never seen it read it fresh.</p><p>That’s not laziness. That’s how you ship something you’re not embarrassed by two days later.</p><p><em>I write about building AI products in public, PM career stories, and the India tech job market. Follow me on Medium if any of that is useful to you.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2f941104a615" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/i-ran-8-ai-agents-overnight-on-my-own-startup-they-found-50-bugs-i-missed-2f941104a615">I Ran 8 AI Agents Overnight on My Own Startup. They Found 50 Bugs I Missed.</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Recruiters Rejected My Resume. So I Fed It to an AI 300 Times and Fixed It.]]></title>
            <link>https://debashisnayak.medium.com/recruiters-rejected-my-resume-so-i-fed-it-to-an-ai-300-times-and-fixed-it-c6737eae8e0a?source=rss-d5ee2eb07331------2</link>
            <guid isPermaLink="false">https://medium.com/p/c6737eae8e0a</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[career-advice]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[job-search]]></category>
            <category><![CDATA[resume]]></category>
            <dc:creator><![CDATA[Debashis Nayak]]></dc:creator>
            <pubDate>Tue, 14 Apr 2026 18:03:14 GMT</pubDate>
            <atom:updated>2026-04-14T18:03:14.846Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c4P-jrntgydgAoQY6phmfg.jpeg" /></figure><p><strong>What 300 AI resume roasts taught me about why PMs get screened out before a human ever reads their application</strong>.</p><p>Nine months of job hunting. Forty companies researched. Nine applications sent in the first wave. Two first-round interviews. Zero offers.</p><p>My resume wasn’t the only problem. But it was a problem I could actually measure.</p><p>So I built a tool that scores resumes against job descriptions - tells you exactly which keywords you’re missing, which phrases you’re using wrong, and what score an ATS would assign before a recruiter ever sees your name. Then I ran my own resume through it 300 times, against every type of PM role I was targeting.</p><p>What came back was uncomfortable. My resume was performing well against generic product manager listings. It was failing, consistently, against every fintech and growth PM role I actually wanted.</p><h3>The problem with resume advice: it’s mostly guesswork</h3><p>Most resume advice is written by people who’ve never looked at match data. They’ve reviewed resumes as hiring managers. They’ve noticed what looks good on paper. That’s useful - but it’s not the same as knowing how an applicant tracking system scores a document before it reaches a human.</p><p>ATS systems don’t read resumes like a person does. They parse keyword presence, calculate weighted match scores against the job description, and filter below a threshold. The threshold varies by company, role, and how many applications came in that week. You don’t know the cutoff. You don’t know the weighting. You just know whether you got the call.</p><p>Running 300 scored match reports gave me something most job seekers never have: actual data on where the gaps were.</p><h3>What the scores told me</h3><p>My baseline resume scored 71–78% against generic PM roles - solid. Against the fintech and growth PM roles I was actually applying to, it dropped to 52–61%. Below most ATS cutoffs.</p><p>Three patterns were consistently killing my scores.</p><p><strong>Pattern 1: Metrics present, framing wrong</strong></p><p>I had strong numbers. Doubled auto insurance revenue from $120K to $350K per month in six months. That’s a real result. <strong>The problem:</strong> I’d written it as a product achievement. Fintech roles are scanning for growth PM language - “<strong>MoM revenue growth</strong>”, “<strong>ARR expansion</strong>”, “<strong>conversion rate lift</strong>”, “<strong>revenue per user</strong>”. My bullet described what I built. It didn’t speak the language of what I grew.</p><p>The fix wasn’t adding keywords. It was reframing the same achievement in growth PM vocabulary. Same number, different signal.</p><p><strong>Pattern 2: Domain language mismatch</strong></p><p>I’d spent two years in insurance-adjacent fintech. My resume said “<strong>auto insurance vertical</strong>” and “<strong>embedded widgets</strong>.” The JDs I was targeting said “<strong>embedded finance</strong>”, “<strong>partner distribution</strong>”, “<strong>B2B2C channel</strong>”, “<strong>insurance-as-a-service</strong>.” These are the same things. An ATS doesn’t know that.</p><p><strong>The fix:</strong> map your company’s internal vocabulary to the industry vocabulary in the JDs you’re targeting. Every company names things differently. The JD is telling you exactly what words they want to see.</p><p><strong>Pattern 3: Builder language, not outcome language</strong></p><p><strong>“Built CNF partner pages.” “Designed in-app quote flow.” “Launched embedded widgets.”</strong></p><p>These are feature descriptions. Fintech PMs at growth-stage companies aren’t selected for building features. They’re selected for moving metrics. The same experience rewritten: “<strong>Expanded top-of-funnel via partner channel, driving 40% of new insurance activations.</strong>” Now it’s an outcome. ATS scores it differently. A hiring manager reads it differently.</p><h3>AI tailoring vs. keyword stuffing — they are not the same thing</h3><p>The instinct most job seekers have after reading this: add the missing keywords to the resume.</p><p>That’s keyword stuffing. It looks like this: “<strong>Experienced in MoM growth, ARR, embedded finance, B2B2C, and conversion rate optimization.</strong>” It’s a list. It signals you know the words but not the work. Experienced recruiters see through it in three seconds.</p><p>AI tailoring is different. You give the model your raw experience - the real outcomes, the real context - and ask it to reframe how that experience is described using the vocabulary of the target role. You’re not adding things that aren’t there. You’re translating what is there into language the reader expects.</p><p>The distinction matters because one gets you past the ATS and embarrasses you in the interview. The other gets you past the ATS and gives you something to talk about.</p><h3>The system I use now</h3><p>One resume backbone. Infinite targeted variants.</p><p>The backbone is the canonical version of my experience - every achievement written in clear, specific outcome language, with real numbers, in my natural voice. I don’t submit this directly to anything.</p><p>For each application, I run the JD through <strong>Jobtune.io</strong>, get the gap report, and rewrite the three to five highest-gap bullets using AI. The rewrite takes 1–2 minutes. The result is a resume that scores 78–85% against that specific role - above most ATS thresholds - without adding anything that isn’t true.</p><p><strong>I run a final check:</strong> does every rewritten line still describe something I actually did? If yes, it goes in. If no, I find a different way to phrase what I actually did.</p><p>The output isn’t a fabricated resume. It’s the same experience described in language that the right people can actually find.</p><h3>What I’d tell myself nine months ago</h3><p>Don’t send the same resume to 40 companies. Send 40 different versions of the same resume.</p><p>The experience is the same. The metrics are the same. The work is the same. But a growth PM role at a fintech company and a generalist PM role at an enterprise SaaS company are looking for different signals in the same document. Giving them the same document is a choice - just not a good one.</p><p>The tool I built to do this is <strong>Jobtune.io</strong> . It’s live. You can run your own resume through it. You don’t have to do 300 rounds yourself - though if you’re nine months in with two interviews and zero offers, you probably should.</p><p><em>I write about job hunting, building in public, and AI tools for PMs and engineers. Follow me on Medium if any of that lands.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c6737eae8e0a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>