Khavin Shankar

The computer next door, 2008 · 4th grade

My fascination with computers began at a neighbour's second-hand PC. I watched, then played: Age of Empires, gardening simulators and whatever else it had. There was no thought yet that I might someday work with computers. They were just something I wanted to spend more time around.

A year later my cousin got an 8-bit console. Seeing it made me want one of my own, and eventually my mother bought me one. For now, technology was still exactly that: a fascination.

The ₹1 crore promise, 2011 · 6th grade

Around sixth grade, my father made me an unusual promise: when I got married, he would give me ₹1 crore to invest. It was an enormous, almost abstract number. He probably does not remember saying it. I do.

Around the same time, whenever someone asked what I wanted to become, he insisted I say “an investor.” I doubt I understood what that meant. But the word, and that number, stayed somewhere in the back of my mind.

Sixth rank, second-hand PC, 2012 · 7th grade

I was studying away from home in a hostel when I finished sixth in my class. My parents never told me they had planned anything. When they came to pick me up on annual day, they surprised me with the thing I had wanted for years: a computer of my own, second-hand.

I mostly used it for exactly what you would expect from a kid: games, movies, and whatever caught my attention. But for the first time, the computer I had admired from next door was at home, waiting for me.

The numbers that wouldn't leave, 2013 · 8th grade

My father had closed his small jaggery business and moved into lending, a small-scale kind of micro-finance. Explaining it to me, he mentioned that at his rates, money roughly doubled every two and a half years.

I could not stop running the numbers. Two and a half years, double. Another two and a half, double again. Then again. I played with those calculations over and over, watching how quickly they grew. Long before I knew much about finance, compounding had stuck.

First laptop, first "program", 2013 – 14 · 9th grade

In the holidays after eighth grade came my first laptop. We went to Ritchie Street in Chennai and window-shopped through what felt like twenty stores before choosing one.

Not long after, I tried to solve a real problem with it. I had watched my father keep track of his lending, so I built him something in Excel to track the money he lent and automate the calculations around it. It took me about two days. Today it would take five minutes.

That is exactly why I remember it. It was not sophisticated and I barely knew what I was doing. The interesting part was the instinct: I saw a real process and wondered whether a computer could make it easier.

Leaving Nellore for Kota, 2015

After 10th grade I left Nellore for IIT coaching in Kota, a place known for intense preparation.

My father had once dreamed of studying at an IIT himself, but debt meant he never had the chance. He wanted me to have the one he did not. So Kota became the next stop.

Escape into anime, Late 2015

Kota was hard in ways I had not expected. I was the only Tamil student in my institute, the language was a barrier, and the coursework was relentless. Along with a new kind of freedom, I began to drift.

My escape became anime. It started with a few skipped classes. Then a few more. Eventually there were days I skipped everything and stayed in, watching, avoiding a workload that felt like it was getting away from me.

100 bits, Mid 2016

Somewhere in the middle of all that drifting, I stumbled onto something completely different. On my first smartphone I found Zebpay, which was giving away 100 bits of Bitcoin for signing up. That was worth about ₹4.

₹4 was not exciting. 100 bits somehow was. I signed up. Then I got my friends to sign up, then my parents. I kept collecting those tiny amounts and watching the number grow. I had found Bitcoin through a signup reward, but something about accumulating it had caught hold of me.

Half a Bitcoin, gone in a week, Late 2016

By late 2016 I had about half a Bitcoin, roughly ₹40,000 then. It came together in pieces: crypto bounty programs, money saved from lunch, and money I had asked my parents for under the pretext of tuition that did not exist.

Then I found Richmond Berks, a platform promising around 1% returns every day. I put the entire half Bitcoin into it. Within about a week, the platform disappeared, and everything with it.

I kept the failure to myself. It was my first expensive encounter with something I had barely understood until then.

What I learned: A scam can look convincing right up until the moment the money disappears.

A second chance in crypto, Early 2017

Strangely, losing half a Bitcoin did not kill my fascination with it. The price was beginning to rise and I wanted another chance. This time I did not have the money.

So I went to my father, told him what I wanted to do, and asked him for a loan. He trusted me with ₹1.75 lakh, an enormous amount for me then. With it, I bought one Bitcoin.

The IIT setback, Mid 2017

Eventually life caught up with me, and my two years in Kota came to an end. I did not make it into IIT.

For me it was not entirely unexpected. I knew how I had spent those two years, and I knew I had not made the most of the chance I had been given. For my parents it was a shock. I packed up and came home.

At the time, not making it into IIT felt like the setback. A week later, my 12th-standard results arrived.

4, 7 and 9, Mid 2017

I had not merely fallen short. I had failed by a huge margin: 4 in Mathematics, 7 in Chemistry, 9 in Physics. Suddenly IIT was not the problem any more. I was not even eligible for a B.Tech.

We went from one tuition centre to the next looking for a way forward. Some said engineering simply was not realistic for me, and that I should try a polytechnic or a diploma instead.

My father was not ready to accept that. After Kota, after IIT, after those marks, he still trusted me. That trust changed how I responded to failure. For the first time in a while I started seriously looking for a way forward, and we found NIOS, the National Institute of Open Schooling: another route to finish school and become eligible for engineering.

What I learned: The obvious path had closed. There was still another one.

Everyone else moved forward, Late 2017 – mid 2018

My friends went off to college. I stayed home. One of them, Rahul, had made it into IIT Madras. We had not really kept in touch, but one day I messaged him, and he mentioned they were learning C in their first semester.

That gave me a simple thought: if that is what they were learning at IIT, nothing stopped me learning it at home. I found Bucky Roberts’ tutorials on YouTube and started C++. Unlike the Excel sheet years earlier, this was my first deliberate attempt to learn to program.

And once I could program, an old curiosity came back. This time I went beneath Bitcoin’s price, learning how blockchains actually work and building simple blockchains and smart contracts of my own. My friends had moved forward while I stayed behind. But in that year, I started moving in a direction of my own.

Finding a way into engineering, Early 2018

I finished 12th through NIOS with 59%, which meant IIT was out of reach. I still wrote JEE Mains, and once again I did not make the cutoff.

So I looked elsewhere. I wrote every entrance exam I could and explored whatever options were still open. One of them was SRM.

A year earlier I had come home from Kota, failed 12th, and been told that a B.Tech might not be for me. Now I had a way into engineering.

Finding my footing, 2018 – 19

I entered SRM with a quiet resolve. For the first semester I kept things simple: do the coursework, make a few friends, find my footing again.

Then the results came out. 10 CGPA. After Kota, IIT and failing 12th, I had not expected it. That number gave me back some of the confidence I had lost, and more importantly the confidence to start exploring again, this time in a better direction.

I explored investing, space, psychology, anything that caught my attention. Preston Pysh's investing course stood out: it taught me to read a company through its fundamentals and to see a stock as a business, not a price moving on a screen. I would return to those ideas years later. An NPTEL course on human behaviour was part of the same wide-open year.

Standing taller, 2019 – 20

By second year I had found my rhythm: friends, a place among the class toppers, and a reputation for being good at programming. Very few people knew I had failed 12th two years earlier.

My learning went deeper. Andrei Neagoie's web development course led to his React course, and soon to a habit of hunting down interesting courses and binging them one after another.

My professor, Vijil, asked me to help with a research paper she was writing. I helped organise the material, rephrase sections and shape it into a paper. I was not an author, but I learned how papers are built, which mattered when my own research came along. Years of anime had, unexpectedly, made my English good enough to be useful.

Jardiniere: friends are not a team, Oct 2019

My first idea I wanted to turn into a business: a crowdfunding platform where startups could offer equity for funding.

It failed almost as soon as it started, for two reasons. I pulled in nine friends, handed out roles and set off, without seeing that they neither shared the vision nor were really interested. And I did not have the technical skill to build it alone yet, so it went still and silent.

What I learned: Recruit only people who share the vision, and keep moving, even in small steps, while you learn.

Bloody Mary, at a hackathon, 11 – 13 Jan 2020

Bloody Mary connected hospitals that need blood with registered donors nearby. Donors signed up with their blood group. When a hospital posted a request, the app found matching donors in the same city and messaged them on WhatsApp to reply yes or no, collecting volunteers until the requested amount was covered, then handing the hospital their numbers.

At the hackathon, the interface was little more than a form. The whole idea was dismissed as basic and not well thought out.

What I learned: Present with confidence even when the solution is simple. The interface makes the first impression.

From a NAND gate to a computer, The COVID break

Then COVID struck. What was meant to be a short break from college stretched on and on. I went home, but the learning continued: web development, React, and things far outside the curriculum.

One stood out: Nand to Tetris. It starts with a single NAND gate and works upward, building the components of a computer, then the computer itself, then the software stack and operating system that run on it.

I had mostly been learning how to program computers. Seeing how one is built from the ground up was a completely different way of looking at them.

A redesign nobody asked for, 2 – 8 Jun 2020

Our first try at freelancing. Stuck at home in COVID, Ashwin and I wanted to do something, and his brother, a distribution head at Asian Beverages, suggested we redesign the company's landing page.

We jumped straight in, without checking with the people who ran the company whether they wanted it redesigned at all, or why. Our homepage was better, if I say so myself. We never managed to pitch it, or even get it in front of management.

What I learned: Understand the requirement from the right people first. A better-looking page may not be what the business needs.

Ideas in the margins, College, undated

Like any aspiring entrepreneur, we had a bucketful of ideas that never reached pen and paper: a botnet, predicting crop demand with GIS mapping, hacking satellites, an animated movie, an anime recommendation engine.

None of them were built. But each one sent me into a new corner of technology. Like most software engineers I was curious about hacking, so I watched videos and took Udemy courses to understand it better, and that is where the botnet idea came from. The satellite curiosity grew into a real project, Cubesat.

These ideas did not ship. They are more like a record of what I was learning.

Google Developer Student Club, Oct 2020 – Jul 2021

When the Google Developer Student Club started at my college, a friend who headed it invited me in for my technical strengths. I hoped it might open doors toward entrepreneurship or deeper learning.

It turned out to be more passive than that. I helped with websites and virtual events. Not a big milestone, but another path I tried.

A+1, the game we built, Nov 2020 – Jan 2021

Like all programmers, we decided to build our own game. This time I was deliberate about the team: we picked six genuinely interested people and split the work into characters, terrain, mechanics, networking and animation. Each of us learned what our part needed: Unity, Maya and all the tooling around them.

We started building. Then the disagreements began, about the character design and the direction of the game. What looked like a small creative argument slowly chipped away at our motivation. Momentum faded, progress stopped, and the project went quiet.

We had solved the problem from Jardiniere: everyone genuinely wanted to build something. We had not solved alignment. We all wanted to build a game. We did not share the same picture of the game.

What I learned: Interest gets people into the room. A shared vision keeps them there.

24 of 50,000, Jan – Mar 2021

In early 2021 I was selected for the CoronaSafe Engineering Fellowship, an AICTE and Pupilfirst initiative to find and train student engineers to build open-source software for public service. I was one of 24 chosen from around 50,000 applicants.

It began with about 120 hours of hands-on training in Ruby, Rails, PostgreSQL, React and ReScript. But the important part came after: instead of stopping at coursework, we were expected to apply it on real open-source projects that people would actually use.

Three days at Anand Techverse, Jan 2021

Like almost every college student, I was hunting for internships wherever I could find them. Through a friend's father we found one at Anand Techverse and joined, expecting a look at the industry.

The sessions began with the very basics of computers and operating systems, things we were already comfortable with. After three days we realised we were not learning anything new and stopped going. They still gave us the certificates.

Cubesat: a satellite on paper, Jan – Apr 2021

The most cross-domain project we ever did. During the COVID break, Ashwin and I fell into space, found a video about intercepting satellite broadcasts, and somehow progressed to wanting to build a satellite.

We learned to program microcontrollers, design PCBs, solder, pick sensors, manage power, and read up on weather balloons, materials and anything else a cubesat needs. But it stayed a design. Money was the constraint, and then the break ended and job hunting began.

What I learned: Building a satellite was reason enough for us. We could explain how to build it, not why it would be useful to anyone.

Learning in the second wave, Mar – Dec 2021

My first commit in this story is a login page for Arike, a Rails app palliative-care teams use to plan home visits, on 17 March 2021. By December I had rewritten its agenda in ReScript and React.

In April, during India's second COVID wave, I worked on Life, a directory of oxygen, ambulances, hospitals and helplines by state and district. Then came CARE, the open-source hospital management system. My first pull request there was a multi-select filter on 24 April; by the end of the year I had opened more than a hundred, almost all merged, and fixed more than 80 issues.

This was software beyond my own projects: an existing codebase, other developers, being reviewed and reviewing, and fixing the unglamorous things real software needs.

A drone that follows a car, Apr – Oct 2021

Our faculty advisor suggested a patrol drone to a friend, who was overwhelmed and bored by it within days. I liked the idea, and his quitting pushed me to try it with Ashwin.

First the material: ROS, Unity simulation, AirSim and YOLO. We were slow. We were used to learning from YouTube, and there was very little of it on this. Just setting up AirSim took a week or two. We worked on and off until we finally had a simulation of a drone following a car.

By then the hackathon was long gone and the project was shelved. Not a total loss: friends later presented it as their minor project, and it impressed their teacher.

What can we cook?, 26 Jun – 7 Jul 2021

A friend suggested we build something together, and we landed on a simple idea: point a camera at what's in the kitchen, recognise the fruit and vegetables, and suggest recipes you can actually make.

We started collecting data and, as usual, I started filling my gaps: YOLO, computer vision, NLP. But our commitment to building it was not the same. He had felt left out and wanted to build something, and that was not reason enough to carry it. As the imbalance grew clearer, my motivation faded too.

What I learned: Don't just ask whether someone is interested. Understand what that interest is rooted in.

A paper at IEEE, Jul 2021 – Jan 2022

The MIT Square Indo-Global Research Internship picked five three-person teams from our college to research an idea of their own under an experienced professor.

Our question: how to find the most influential nodes in a graph with geometric deep learning. We were guided by Dr. Amira Alturki of the University of Business & Technology, Saudi Arabia, and Dr. Mithileysh Sathiyanarayanan, CEO of MIT Square.

We published "Predicting Nodes for Effectively Spreading Vaccine Awareness: An Influence Maximization Approach" at the 4th IEEE ISAECT conference in 2021.

Learning to teach at ProGrad, Sep 2021 – May 2022

ProGrad trains new hires for IT companies. As an Expert Mentor Developer intern I fixed bugs in their software, but mostly I taught.

It began with two months of Java, TDD, Spring Boot and GCP, with demo sessions of what we learned. Then, with two other mentors, I took a batch of 24 mentees from UST Global, eight of them under me. After that, a batch of 12 from Taylor & Francis on Git, HTML, CSS, JS and the MEAN stack.

It sharpened my technical knowledge, but it changed my soft skills more: communicating, delivering content, and being far less afraid of the stage.

The cheaper bid lost, 21 Sep 2021

Through a friend's father we met the owner of PEPS mattress, who wanted to take his business online. He invited another local company to quote as well.

We specced out better features for far less: a custom e-commerce site for ₹10–20K. The other company quoted five times as much, and won.

We had focused almost entirely on technical value. We had not thought about how a business judges the people behind a proposal. An established company signals continuity, accountability and lower risk in ways two students saying “we can build this” cannot.

What I learned: Before people can value what you build, they need a reason to trust you to build it.

Into 10BedICU, Nov 2021 – Apr 2022

CARE opened another door. Srikanth, one of eGov Foundation's founding members, had started 10BedICU: ten-bed ICUs in rural hospitals, linked to larger ones so specialists can support doctors remotely. CARE was a natural fit as the software for those TeleICU workflows.

So the fellowship continued as an eGov-sponsored extension. I worked on CARE's critical-care and TeleICU modules: how consultation data is collected, and screens for monitoring ICU patients remotely. Much of the rest was turning a project into production software: bugs, warnings, small features, clean-up, deployment.

It was one of the first times software stopped feeling like something I was learning to build, and started feeling like something I could use on problems that mattered.

From filters to features, 10 May 2022

I joined eGov full time on 10 May 2022, and my changes got bigger. I rebuilt CARE’s notifications panel and fixed web push underneath it, added discharge reasons end to end, swapped the location picker for Google Maps, and added drag-and-drop to the shifting and resource boards.

The largest piece was coded diagnoses. I wrote a scraper that walks the WHO's ICD-11 browser and produced a local copy of the whole classification, then wired ICD-11 diagnoses into the consultation form.

I also joined the contributor leaderboard two weeks after it started: profiles, search, collaboration achievements and stale-PR detection. The activity data its scraper collects is what my eGov story is built from.

Half the pay, the right room, 18 Jul – 23 Sep 2022

In final year I had taken a pre-placement offer from rtCamp, starting six months later: WordPress and PHP, weekly course material, a coding assessment before each deadline, then a rapid-fire technical round. Late nights became the norm. The rounds pushed me, but the sharp criticism after them, on top of endless deadlines, made learning feel like college again: chasing scores.

A month before it began, I had sat in an eGov meeting with the 10BedICU team. They were welcoming, there was freedom to explore modern technology, and I was working alongside people like Srikanth, who had been Aadhaar's CTO.

Two months into a six-month internship, I left rtCamp for eGov, at roughly half the pay.

What I learned: The right environment, people and freedom to explore change how much I want to learn and build.

Into the ICU, Aug 2022 – Jan 2023

10BedICU runs ICUs in rural government hospitals remotely: a specialist watches the bed through a PTZ camera, and the bedside monitor streams vitals over HL7. Between CARE and that hardware sits the TeleICU middleware, a Node service at each hospital that talks to cameras and receives the monitors’ readings.

I merged Aakash's automated daily-rounds job, which files a monitor's readings into CARE, with my first attempt at spotting stale readings. I fixed camera feeds and presets, and moved the middleware address from each camera to the facility so a hospital's devices share one endpoint.

Around it: pressure-sore charting, procedures in the consultation form, live vitals waveforms, and oxygen burn-rate estimates for the dashboard.

Teaching again, at Faceprep, Dec 2022 – Jul 2023

I had declined Faceprep's full-time offer in May to join eGov. In December they came back, short-staffed, looking for part-time help. I proposed an hourly arrangement, took one batch, and three months later a second.

ProGrad, its offshoot, had already taught me how to teach, so it flowed: the MERN stack from HTML, CSS and JavaScript to React, Express, Node and MongoDB. Both batches went well. They offered a full-time role again, and I declined again; I was all in at eGov.

A side gig, but it paid, and it let me keep doing something I love: teaching.

Biller: reality moved faster, 15 – 16 Dec 2022

Ashwin's father was opening a shop and needed billing software, and Tally felt expensive for what he needed. We thought we could build something simpler.

We started a basic GST billing app: products, customers, quantities, taxes, invoices. But the shop was already opening, and his father understandably could not wait for us. He bought Tally, the problem disappeared, and Biller with it.

One of our shortest experiments, started and abandoned almost at once, because reality moved faster than we could build.

Plugging CARE into India's health stack, Jan – Jul 2023

In 2023 CARE started talking to national systems. I wrote the Health Claims Exchange (HCX) integration in February: coverage checks, pre-authorisation and claims, each a FHIR bundle encrypted for the insurer. It shipped behind a feature flag in March and later gained communications, so an insurer can query a claim and the hospital can answer with documents.

I also took ABDM, India's rails for health IDs and consented records, through its first two milestones: creating and linking ABHA numbers, and making CARE a provider that sends a patient's records, encrypted end to end, when they consent. M1 and M2 merged on 21 July 2023 with a 3,664-line backend PR.

hTab: AI in a new tab, 21 – 29 Jan 2023

I spent most of my day in a browser, so why couldn’t the new-tab page be a workspace? I built a home screen with a clock, quick notes stored locally, and a search bar.

I was experimenting with OpenAI's APIs at the time, so I turned the search bar into an AI prompt on text-davinci-003: type a question into a new tab, get an answer.

It was handy while I taught at Faceprep, and I used it as an example for students to build their own. It stayed a personal tool, not a product.

The other side of Pupilfirst, Feb – May 2023

The Pupilfirst team knew me from the CoronaSafe fellowship and invited me back as a teaching assistant.

I answered student questions on the forum and Slack, and reviewed submissions, from level assessments to capstone projects. It sat naturally on my web development background and ran smoothly for months.

Towards the end I was juggling eGov and several other commitments, and I stepped back because I could not give it the focus it deserved.

One project became three, Mar 2023 – Feb 2024

IT-Metrics came to me through an agency: a p5.js project with Fabrice and Guillaume called Cryptopics, which turned Ethereum blockchain data into generative art. A contained frontend job took me into Ethereum data, BigQuery, Solidity and the practical problems of processing a lot of it.

Fabrice then brought me onto doku.host as the sole developer under him and Guillaume: a platform for sharing PDFs, documents, published content and whole HTML sites, with authentication, storage and subscriptions. By February 2024 v1 was in production, and I had started the foundations of a third product, Goverly.

Almost a year, three products, and growing independence. It became one of my longest freelance relationships.

The tracker I didn't keep, 16 – 23 Apr 2023

By now money was arriving from several projects, from different sources, on different schedules. I wanted one place to see those streams next to my recurring expenses, so I built a Cash Flow Tracker: streams, groups, transactions and charts over time.

I used it for a while, but never built the habit of keeping it up to date. When I stopped using it, development stopped too.

No failure or pivot behind this one. I had a problem, built something for it, and found I did not care enough about the problem to keep maintaining the answer.

Teaching the middleware to see, Jul – Dec 2023

Back at the bedside. First, device health: camera and monitor uptime sampled every minute, so we could see which devices were actually online.

A colleague had started reading vitals off monitor screens with OCR. I got it working end to end: which camera preset points at which bed's monitor, cameras locked while a snapshot is taken, and every reading uploaded with the image it came from.

For ABDM I mapped facilities to the national registry and built hi-profiles, a React library that renders ABDM's health records, ready for the day CARE would receive them. 2023 was also my heaviest year of reviewing, 1,026 reviews, and in December I started running CARE's weekly releases.

Ciragu: ambition outran us, Jul 2023 – May 2024

A year after college, a hackathon reminded me of our old drone project, and I asked Ashwin to pick it back up, far more ambitiously: a modular operating architecture for autonomous drones, where perception, navigation, flight control and obstacle avoidance are replaceable modules, separating the drone’s brain from its airframe. Ashwin named it Ciragu.

We explored AirSim, Unreal, ROS, SLAM, behaviour trees and reinforcement learning, then aimed at a government defence programme: train behaviour on cheap drones, transfer it to expensive ones.

The ambition kept growing; our execution did not. Ashwin kept studying and building while other commitments pulled me away, and I became one of the reasons we stalled. After nearly a year, we shelved it.

When to stop saying yes, Nov 2023 – Nov 2024

Nxerra began with MFund, a community health fund platform, and kept growing: Aira, Healthspan, Phone Darlings, Evostack, Atlas. Aira, an AI hiring platform, became almost a second full-time job, working directly with Isaac. I was even taken to Bangkok to meet him.

But I was paid only when the person who brought me the work got paid by his clients, so payments stayed pending for long stretches. I knew this early, and kept saying yes anyway. When payments became a problem, my motivation and availability suffered, and the clients saw missed deadlines, not the dispute behind them.

After a year of the same cycle, I stopped. My rule since: no long-term client relationship entirely through a broker, and never a payment problem the client cannot see.

What I learned: Nxerra gave me some of my biggest projects. It also taught me when to stop saying yes.

Roosty, and meeting Wei, Dec 2023 – Feb 2024

I met Wei through an agency. He was building Roosty, an Australian marketplace for parking near stadiums on game days, and I came in to help with Supabase. Within weeks the team had changed around me and I was working across most of it: maps, listings, bookings, payments, payouts, auth, emails and deployment. By late January Wei had moved me onto his next project, Propflow.

The trouble was the arrangement. I was not being paid what I had agreed with the person who brought me in, and I handled it badly: I stepped away without explaining why. Wei only saw the work slowing down.

The relationship survived anyway. More than a year later, Wei remembered our work, and that led to Funky Food and Loopin.

What I learned: A payment problem with a broker should never silently become a trust problem with the client.

Good pay, wrong work, Dec 2023 – Feb 2024

Another agency contract, at ₹500 an hour, which felt like good money. It was pitched as WordPress plugin and theme development. In practice it was maintaining other businesses’ sites: content updates, small styling changes, routine requests. The client was harsh and at times rude.

Three weeks in, I asked to leave. It took persistence to be taken seriously; I stayed two more weeks until a replacement was found, handed over, and moved on.

About six weeks in all, and it taught me something about myself: I am happiest with something to build, solve or think through.

What I learned: Good pay alone doesn't make a project worth taking.

Consent, and vitals read by a vision model, Jan – Jul 2024

ABDM's third milestone flips CARE's role: a hospital can ask a patient for consent and pull their records from anywhere in the network. I opened the backend in January and it merged in May, with Aakash and Gigin.

In the middleware I moved vitals OCR onto a vision model. Every hour it checks each monitor; if the monitor’s own readings are missing or stale, it points the camera at the screen, asks a GPT vision model for the readings as JSON, sanity-checks them, and files them in CARE. It also measured itself, comparing each reading against the monitor’s own value at the same moment.

I ran many of the weekly releases, and in April added a workflow that tags and drafts each one.

A payment problem became a trust problem, 2 Apr 2024

A few months into IT-Metrics, the person handling my engagement offered to continue the same work directly through them, and I agreed. Over time the payments became delayed and inconsistent, and by the end a significant amount was unpaid.

Instead of raising it with Fabrice directly, I simply became less available, and that was the bigger mistake. My frustration was with the intermediary, but Fabrice only saw the work slowing down.

On 2 April 2024, I stepped away from an engagement that had begun more than a year earlier.

What I learned: When an intermediary relationship breaks down, don't let poor communication turn a payment problem into a trust problem with the client.

Breaking up the monolith, Jul – Oct 2024

National integrations like ABDM and HCX lived inside CARE's core, and not every hospital used them. Vignesh had added plug support, separately packaged Django apps a deployment opts into, and I moved ABDM out first: 5,225 lines deleted from core, and a new repository, care_abdm.

Inside the plug I moved to ABDM’s v3 APIs and replaced the Java encryption sidecar with an in-process Python one, so a deployment no longer needs a second service. Then ABHA cards, and Scan and Share: a patient scans a QR code at the counter and their profile arrives in CARE.

HCX followed (3,105 lines out of core). I wrote care_hello, the template for new plugs, and a proposal for scheduling in CARE.

OnQ: a queue that lost its rhythm, Oct 2024 – Mar 2025

I met Gunal at my grandfather's funeral, and somewhere in our conversations we got onto entrepreneurship. He had an idea for salons, Book Your Stylist. What caught me was the queue underneath: a general system for tokens, walk-ins, appointments and notifications that could work almost anywhere, even in a hospital system like CARE.

I built the backend fast in NestJS and Postgres while Gunal handled the frontend and design. Then we lost momentum. I preferred to move quickly and decide as I built; Gunal preferred to align first. As other work took over I became hard to reach, calls moved, wireframes waited, and OnQ quietly went still.

Neither working style was wrong. We never agreed how ours would fit together.

What I learned: A good partnership isn't finding someone who works like you. It is agreeing, early, how you will work together.

Micro-frontends on CARE v3, Oct 2024 – Apr 2025

The frontend followed the backend out of the monolith. Gigin built a plugin system into CARE's frontend; I moved HCX and then ABDM onto it, taking more than 5,000 lines out of core, and rebuilt the ABDM frontend from scratch as a micro-frontend when the system moved to module federation.

At the same time CARE itself was being rebuilt as v3, with clinical records shaped like FHIR resources. My part began with medications, models through to the printable administration record, and moving ABDM onto the new models.

Every place a plugin can attach to core is a deliberate seam, and I added many of them in early 2025, along with the plumbing plugins need.

Quip: the companion we couldn't afford, 11 – 22 Dec 2024

Six months after Ciragu, Ashwin and I had not built anything together. So we booked a stay in Coimbatore for a month and tried a new idea.

Quip was a gaming companion that watches you play and joins in: suggestions when they help, reactions to what happens on screen, or, by default, a snide voice telling you how useless that dodge was. We tried local Whisper for voice, screen capture and OCR, and multimodal models. What survived in code was simpler: an Electron app reading text off the screen.

Economics ended it. To feel real, Quip had to understand the game continuously, and video-capable models were far too expensive then to offer that at a sensible price. It did get us building together again.

Memoir: early, but we couldn't explain it, Feb – Jun 2025

AI chat apps started every conversation with amnesia. Ashwin and I wanted an assistant that remembered, and that community plugins could extend: Tally for sales, IoT for your home, and plugins talking to each other. We built per-user memory with embeddings, a LangGraph agent, web search, journals and real-time voice through LiveKit, and made our first hire, Shadman Khan.

Explaining it was much harder. We pitched to a relative’s company, then a friend’s workplace, five times over, and the idea never landed. Instead of treating that as a validation problem, we reshaped the product around examples people understood, and lost the thesis we started with. We stopped; Shadman’s internship ended with it.

What I learned: Validation is only as useful as the people you seek it from.

Promoted to SDE 2, 1 Apr 2025

On 1 April 2025 I was promoted to SDE 2 at eGov.

By then my work had grown from filters and validations on CARE’s frontend into national integrations, clinical-record models, plugins, reviews and the weekly release train. A milestone inside a much longer stretch of learning with the same team.

Identity at the counter, claims from scratch, May – Oct 2025

I added the other ways ABDM can verify a person at the registration desk: Aadhaar demographics, fingerprints through the desk's biometric device, and in October, face authentication. Encounters, uploaded files and questionnaire answers began registering themselves with ABDM as they were created.

The biggest build of the year was NHCX, the National Health Claims Exchange. I had created the plug in February; in the first ten days of August I built the rest: eligibility checks, claims with FHIR mapping, payer communications, payment notices and insurance-plan catalogues, with the frontend alongside.

Funky Food, briefly, 26 Jun – 18 Jul 2025

On 21 June 2025 Wei and I reconnected, a year and a half after Roosty. A few days later he offered me a role at Funky Food, and I joined on the 26th.

By 12 July a UI problem surfaced: code had been pushed straight to Shopify, causing odd glitches. On the 16th we rolled out new payment-collection logic, aligned to the purchase day rather than the signup day. It failed in production and a few customers slipped through the cracks. My engagement ended on 18 July. The team called it a financial decision; I suspected the timing of those two issues mattered too.

I'm still grateful Wei remembered me after all that time.

What I learned: Cross-team communication, and timing and business logic that agree with each other.

Loopin: owning the product, Sep – Nov 2025

Wei had a working proof of concept for Loopin, an AI voice receptionist for Australian tradies, and on 5 September invited me to build it out, with a possible CTO or co-founder role if it took off.

I reworked the architecture and built the backend and voice systems: telephony wired to the agent, calendars and integrations, and everything that breaks with real calls: latency, failed calls, memory, recordings, interruptions. A receptionist cannot just have a good prompt; it has to know what it needs, when to change direction, and when the call is over. That led me to structured agent flows in LangGraph, and to modifying a LiveKit integration that could not do what we needed.

The pace slowed in late November. Eleven weeks, and my first real taste of technical ownership at the product level.

Payments, and plugins that load at runtime, Oct 2025 – Jan 2026

CARE v3 has billing, so the next plugs were about collecting money. I built a Razorpay plug: a payment link or UPI QR code per invoice, and webhooks that reconcile full and partial payments into CARE automatically.

Plugins had one big limitation: they were compiled into CARE's frontend, so enabling one meant a rebuild. With Gigin I changed that. The frontend now asks the backend which plugins are enabled and loads each one at runtime, inside its own error boundary so a broken plugin can't take the page down.

At the ABDM counter, Scan and Share moved onto CARE's token queues.

Defying a prediction, 23 Dec 2025

As a child, at an exhibition at my uncle's college, a computerised astrology program predicted I would succeed professionally but have only one degree in my life. For whatever reason, that stuck, and I always wanted to prove it wrong.

The other reason was practical. Oozh was taking shape and I wanted it grounded in research, not just engineering. A DBA could give me research relevant to Oozh, C-suite peers, and an academic network. Two mangoes with one stone.

I chose ESGCI, Paris, through UpGrad, and began on 23 December 2025. Personal circumstances meant I could not give it the attention I wanted, so I deferred to September 2026.

State rollouts, and PMJAY on NHCX, Feb – May 2026

2026 brought state deployments. I started a plug for a state HMIS rollout whose first job was billing on booking.

Then a race condition. Scan and Share numbered tokens as count + 1, so two patients scanning together could get the same number. A Redis lock did not fix it: CARE runs each request in a transaction, so my lock was released before the commit. The fix holds the lock until the transaction commits. I would meet that lesson again in August.

The largest piece was PMJAY, the national health insurance scheme, on NHCX: biometric consent, claim workflows, a dispatch state machine, and plan catalogues of up to 100 MB. The backend merged at the end of May.

Now or never, Early Mar 2026

Oozh had been simmering in my mind for over a year and a half. It evolved, expanded and matured on paper: notes, architectures, ideas. But for a long time it stayed exactly that, a concept.

In early March, after a breakup, I gained a new clarity. It became a now-or-never moment, an inner push to stop hesitating. Through April and May I revisited Oozh’s design and refined the architecture, and this time it was not just ideation.

Axia: a good idea, not mine, Mar – May 2026

Axia was the clearest business opportunity Ashwin and I had found. Through his brother-in-law's friend at eZee, a platform bus operators use, came a straightforward suggestion: we build a dynamic pricing engine, they take it to operators.

We built a working prototype: operators define rules on occupancy, departure time, fare and seat type, and Axia recalculates prices, with a simulated GDS, a booking flow and a playground for rules. But the operators’ feedback never really came, and I never felt connected to the problem.

As Oozh took shape, I chose to stop. Ashwin carried Axia further on his own.

What I learned: A promising market is not a reason to build. The problem has to hold me when the opportunity alone no longer does.

JagrIQ: from research to architecture, 20 Apr – 22 May 2026

JagrIQ, a legal-tech startup using AI to make legal knowledge accessible, was exploring edtech: a focused learning platform for law aspirants.

Over about a month I led that initiative: expanding the market research, shaping the product, guiding the design of scalable data schemas and starting structured data collection. We took it to a largely final architecture and were ready to implement when I left to focus on my own startup and the DBA.

It showed me how a small startup team works up close: experiment, decide quickly, and get from research to a grounded product.

Choosing one idea, 22 May 2026

On 22 May I committed fully. I stepped away from Axia, where my focus had been divided, and from my other side projects, and made Oozh my primary pursuit.

From that moment I laid the groundwork, set up the cloud accounts, and began treating Oozh as something I was building rather than endlessly planning. I kept working at eGov for income; my creative energy went into Oozh.

What I learned: I stopped treating every interesting opportunity as one I had to pursue.

Hardening, and one command to run it all, Jun – Sep 2026

At eGov the work turned to reliability. I wrote the integration documents for NHCX and ABDM, and built care_sast for Karnataka’s SAST scheme.

ABDM callbacks got a durable inbox: stored, de-duplicated, acknowledged and processed by workers with retries. Then the transaction lesson for the third time: a worker could pick a task up before its row was committed. The fix is one line, transaction.on_commit.

And a tool for the next person: care_create, a CLI that stands up CARE, its frontend and any plugs locally, and even opens a tunnel so a laptop can receive ABDM callbacks. Others on the team have built on it since.

Back to the DBA, Sep 2026

In September 2026 I picked the DBA back up after deferring it.

The reasons had not changed: research that could give Oozh a foundation beyond engineering, peers with very different business experience, and, if I am honest, a childhood prediction about one degree that I still want to prove wrong.

Oozh, incorporated, 24 Sep 2026

In mid-August I began the incorporation process, and on 24 September 2026 Oozh was officially incorporated.

From there, Oozh became more than an idea. It became the thing I am fully set on building.