01where it started
I grew up on farms. For pretty much my whole life I've lived in a small village in Uttar Pradesh, and electricity came to our village in 2008, when I was 4. Almost everyone around me is into farming, nobody was into coding or computers. I was lucky that my parents were educated, they knew what was important for me and they pushed me in that direction. Most of what I know I learned from YouTube and the internet.
The first time I actually got to know what a computer is, I was in 5th standard (5th grade, so around 10 years old). Our school gave us a computer book and it had all the input devices, the CPU, how stuff moves around inside the machine. I got hooked on it. Funny part is the book was in English and my English back then was not good at all, so I basically learned to read, write and speak English that year just so I could understand that one book.
After that I became kind of unbearable at home. I was literally crying all day so my parents would buy me a computer. In the end one of my uncles arranged an old computer from his office and gave it to me. That was 2016. No internet, nothing, just a box with Windows XP on it. By the end I knew every single feature of Windows XP, im not even joking.
Untill 8th standard I didn't know coding was even a thing. I was mostly playing games, and pretty quickly I started messing with them. Modding GTA Vice City and later GTA 5, changing textures, making the cars go way faster than they should, adding new characters, trying every hack I could find. Installing mods from CurseForge, downloading games from FitGirl and then modding those too. Playing Pokemon on Game Boy emulators. Then Minecraft came and that ate alot of my time too. I didn't know it then but this is where it started.
02finding code
In 8th standard one of my aunts told me about programming languages. Around the same time I used to read the tech section in the newspaper, there was this one section that came with it and I read it with so much interest every time. One day they wrote about an app that teaches Java. I went to Play Store, downloaded it and started learning Java from there.
In 10th my aunt told me about HTML, CSS and the frameworks that were out there. So I started learning those and by the end of 10th I could build a website on my own.
03JEE, YouTube and college
Then came JEE. For anyone outside India, JEE is the entrance exam for the IITs and most of the big engineering colleges, and pretty much everyone around you is preparing for it. I prepared for two years and qualified JEE Advanced in 2023 with a rank of around 16,000 among about 1.2 million candidates. CSE at the IITs closes way before that, so IIT with CSE was not happening, and I took B.Tech in CSE at Rajiv Gandhi Institute of Technology instead (2023 to 2027).
One thing I almost forgot. During 11th and 12th, while I was preparing for JEE, I also got into content making. This was COVID time and me and my younger brother used to sit and break down the content strategies that were making stuff go viral. From there we started two channels, a tech one and a gaming one. I ran the tech channel and my brother ran gaming. We couldn't scale them and I left it in my first year of college. My brother kept going though, he's made a lot of videos for companies like MicroAGI and CodeAnt AI (a YC company), and I still advise him on it. In 2024 I also got the PM Young Achievers Scholarship, which is a national award for academic merit.
In college I got way more into coding. First year I was doing every hackathon I could find. I also did SIH (Smart India Hackathon, the national one run by the Indian government) and made it to the finals in first year itself.
Second year i started winning. Won the hackathon IEEE organized in my college. And then the big one, I helped organize an international hackathon between India and Russia under BRICS. It was a bilateral thing between the two countries and somewhere around 15,000 to 20,000 people participated. I managed the site and the infra for it, and the infra we got from college was honestly very little, so a big part of my job was making sure the site could take that much load on what we had.
I also got the judges on board, founders from YC and a16z backed startups:
- Sahil Dhull, founder and CEO of Kyra (AGI interfaces), CS from IIT Kanpur. His previous startup was backed by a16z Speedrun and he's done HCI and ML research at Adobe Research.
- Shivam Bhatia, founder and CEO of Clinsight AI, earlier co-founder at Lunabill (YC F25), and before that at BharatPe.
- Rounak Adhikary, CEO of ProjectX (YC P26).
- Bishal Karmakar, CTO of ProjectX (YC P26), worked with Founders Inc.
- Sourya Majumder, COO of ProjectX (YC P26), also Founders Inc.
04open source
Also in second year I started doing open source. First was the PEcAn Project, I maintained the frontend they use to show their research results. Then I moved to bigger, more backend and frontend heavy stuff like Plone CMS, where I've been contributing since December 2024.
If you haven't heard of it, Plone is a CMS kind of like WordPress but decoupled, React on the frontend and Python on the backend, with lots of plugins and integrations to customise your site. It's actually really fun software to use. A lot of government organizations in Europe and Canada run on it.
The problem with Plone at that time was documentation. There wasn't enough of it and it was really frustrating for new contributors, they kept getting stuck and had to keep pinging maintainers because most of the knowledge was just inside the maintainers heads. The maintainers did create docs issues but very few people picked them up, because to document something properly you need someone who knows the repo deeply to guide you.
So I joined their documentation sprints. We'd take one module per sprint and finish it. I'd be online with the maintainers, mostly Steve Piercy and Victor Fernandez, they'd tell me the correct way something works in Plone and I'd go document it. All the docs are written in Markdown and built into the docs site with Sphinx, there's a toctree that decides how all the pages show up on the site. I completed the documentation for plone.conf this way and did 4 sprints in total, fixing bugs and validation edge cases along with the docs.
It wasn't only docs though. I also migrated the Authomatic library (the auth library Plone uses for social logins) from OAuth 1.0 to OAuth 2.0 for the Twitter/X provider.
05AfterQuery
In August 2025 I started doing freelance work. First was with AfterQuery, a YC startup doing data labeling, remote, for about one and a half months (Aug to Sep). I built debugging and DevOps tasks using Docker with deterministic test suites, aligned with Terminal-Bench validation, and set up automated test environments and scripting workflows for them.
06KD, the legal AI startup
After that I was scrolling tech twitter (I'm on there alot) and a lawyer posted about the problems in their workflows. Lawyers spend most of their day on language work, researching cases and drafting, and by then the reasoning in LLMs had gotten pretty good. So I thought a product could be built here that helps lawyers with research and drafting. That's how KD started, right after AfterQuery.
The first question I asked myself was why can't a lawyer just use ChatGPT. Turns out there are a lot of reasons.
- ChatGPT isn't built for law. You need exact citations from cases, exact citations from the law, and the code sections have to be quoted verbatim. Close enough doesn't work.
- It's hard for an organization to use it. Take a corporate law firm, they only need corporate law, they don't need criminal law sitting in there. So we'd chunk only the data that firm or that team needs, and connect more databases if they need them later. If a firm does both corporate and criminal, every lawyer gets a role based account, data is stored at the organization level, and the agent a criminal lawyer is using only references criminal law.
- Collaboration. There are alternatives to Microsoft Word out there but corporates don't move to them because they don't give organization level and role based access for employees. That's a big reason Word is still relevant. Law firms have different teams working on the same problems all the time, so this mattered alot.
- Repetitive work. Offer letters, term sheets, terms and conditions, the same documents again and again, very boring to do. LLMs can do most of it as long as there's a human in the loop. Same as coding with Claude Code honestly, you still have to look at the output and check if it matches what you meant.
On the tech side KD was Next.js 16 and Express with Drizzle and PostgreSQL. You could upload legal documents (PDF or DOCX), organize them and chat with them, and every user could pick which model they wanted, Claude, Gemini or GPT, all with tool use. I built structured legal workflows like CP checklists and summaries of credit and shareholder agreements that come out as Word documents with tracked changes. For case law I integrated Indian Kanoon and eCourts India, so you could search cases, pull judgments and get court orders analysed by the AI.
Somewhere along the way I noticed my feature list was turning into a copy of Harvey. I wasn't copying them on purpose, it just ended up that way because the problems lawyers have in the US and in India are pretty much the same. But that also made me realize I was building what I thought lawyers needed and not what they actually needed.
So I moved to Delhi for 3 months. I spent basically my whole day talking to law firms, understanding their problems and changing the product based on that. I piloted with around 10 law firms in Delhi and they became my customers, some of them paid too. Around the same time I went through YC Startup School with KD and had a few early conversations with investors, Lightspeed and DayZero Ventures among them.
But the problems with the business also got clearer.
- Sales cycles are really long. A firm takes 5 to 6 months to close a deal after testing the product thoroughly.
- Data. Some firms were very particular about how we handle their data and would only buy if we deployed it on their own premises and maintained the infra and the software on their own machines, because they've signed contracts with their clients that the data is not shared at any cost.
- Too many competitors. Building the product was the easy part, distribution was the hard part, and I didn't have anyone to handle operations and GTM.
So after working on it for around 6-7 months I shut it down. It was a really tough call but sometimes you have to make those.
Also in 2025, through Y Combinator, I got about $25K in cloud and AI credits across AWS, GCP, Azure, Anthropic and OpenAI.
07Agent Yap, the agentic AI learning platform
After that I had about a month free. I had been studying agentic AI for the past 2-3 months, memory, RAG, all of it, and I kept noticing there was no single place that teaches it properly. YouTube has pretty much everything but its all scattered, there was not one course that takes you through everything in an organized way.
I had a ton of notes and markdown files from my own studying, so I gave them to Claude Code and asked it to turn them into a proper website people can learn from. That became Agent Yap (agent-yap on my GitHub).
The first version looked really bad. I got kind of obsessed with the frontend after that because I wanted it to feel nice for whoever comes to read and study on it. I iterated alot using tools like Paper (a newer, better take on Figma) together with Claude Code and ended up with a minimal look. Now it's quite detailed. It has practice sections, animations for the stuff thats hard to get from just text, and code snippets from well known open source repos, like if I'm explaining the agentic loop I show how a coding harness like pi actually does it.
Right now it's 199 slides across 18 chapters, all pre-rendered so it loads fast, with BM25 full text search and keyboard or swipe navigation. There's a Q&A thing built on Claude with RAG that answers your question and links you to the exact slide the answer came from, and a Gemini chat panel on every slide where you bring your own key. Built with Next.js 16, React 19, TypeScript and Tailwind v4.
08QuicShop
Since June 2026 I'm a founding engineer at QuicShop. QuicShop is a store builder, kind of like Shopify but AI native and built for Indian sellers.
My first day was funny. The team was going to a startup event in Bangalore where a lot of sellers were going to be, and the founders asked me to go around the stalls and pitch our product. It was literally my first day. So I learned the product on the spot and kept my approach simple: ask the seller what problems they're facing in their business right now, then steer the conversation towards how QuicShop could help with that. Sometimes what they said had nothing to do with what we sell and I had to think on my feet and somehow connect it back. By the end of the day I had closed around 30 users for our pilot, and some of them bought the paid plan too. I was honestly shocked myself, I didn't know I could sell.
The founders thought I should move to the marketing team after that but I wanted to work on the technical side, so I asked them to keep me in engineering and I got backend and frontend work.
Some of what I've done here so far:
Payments. We had no way for customers to pay us for a subscription, everyone was just using it for free. I built the whole subscription and recurring billing system on Razorpay so sellers could subscribe and get access. Payments is where the edge cases hurt, retries, webhooks, failed charges, partial charges, so alot of the work was handling those properly. And since QuicShop is multi-tenant, everything is scoped to the shop it belongs to.
Getting sellers on board. Onboarding new sellers was hard. The ones who already have a store are mostly on Shopify and don't want to move their whole store. So I went after the ones who don't have a storefront yet, people selling through Instagram with a Google Form link or a WhatsApp number in their bio. I built an n8n automation that scrapes those profiles, adds them to the open source CRM we host on our own droplet, and drafts a personalised WhatsApp DM for each seller based on what's in their bio. Someone from the marketing team checks the draft, edits it if needed, and sends it. For people who don't reply we call them manually for now, and I'm looking at a voice AI agent for those calls so we can reach more sellers for less money.
Inventory and shipping. Right now I'm shipping real-time inventory sync across D2C and B2B channels, and I've integrated Shiprocket for shipping and WhatsApp automation.
Faster checkout. Customers had to type their whole address into a text box at checkout which is a really bad experience. Saving the address the first time helped for repeat buyers, but we wanted people to get the full address suggested while they're typing. So I added Google's address autocomplete. It's normally paid but Google has a startup program where you get it free for 2 years if you apply, so we got it at no cost.
CRM costs. We were paying for Salesforce at around $40 per seat, and with a team of 20-25 that came to roughly $1,000 a month. That felt way too expensive. I went through the open source CRMs out there, estimated what each would cost to host on a rented VM, did some back of the envelope math on how much compute our team would need, and picked Twenty because it covered the features our marketing team actually needed. It runs on a small DigitalOcean droplet for about $50 a month. Since it's open source the maintainers take care of the software and I only maintain the infra.
09BrainStorming, the side thing
On the side I'm building BrainStorming. One LLM on its own isn't great at divergent thinking, it kind of settles on one idea fast. So this is like a panel of experts. Claude Opus does the deep research first, then fans out to Codex and Cursor CLI models (Grok, GPT, Gemini) in parallel. Their answers get mixed into blended ideas, and then it converges on a scored, red-teamed verdict, all tracked on a shared ledger where better outputs carry more weight.
10how I work
If you read all of this you'll probably notice I'm all over the place. YouTube, hackathons, open source, legal tech, an education site, sales, payments, growth automation, infra costs. I know it looks weird. But the way I see it, my job is problem solving.
I like working closely with founders and owning things. Founders are already stressed about a lot of problems in their startup, and most of the time they don't have the time to assign you tasks or even explain every problem to you. So I try to find the problems myself. Nobody asked me to work on user acquisition at QuicShop, I picked it up because it needed solving. I like picking up problems I find interesting and getting them done.