The model is rarely the hard part of an AI product anymore; building has never been easier. The real work is judgment: choosing what to build at all, then wringing the most intelligence out of every token it costs to run. Five years in, that judgment is the thing I trust most. One problem is where I really felt it: PS AnalytIQ, the counterfeit-detection product I founded at PharmaSecure. The data had been hinting at fakes for months, one real code scanned thousands of times, and the win wasn't the algorithm so much as deciding exactly what to build around it. It runs in production now, catching fakes before they reach a patient, and it's the same instinct I bring to leading Product at uEngage today.
None of it started in an office. It started on a roller-hockey rink, where I spent more seasons falling than winning before I ever learned to lead a team to a national medal. I never really left it. I still play, I coach now too, and the kids I train have started bringing home national titles of their own. On weekends, I teach children their first lessons in a classroom. Sport and teaching gave me what no framework did: how to read a moment fast, commit before I'm certain, and carry people through the stretch when nothing works. The domains keep changing. The job never does.
It didn't start with job titles.
Made the team ship roughly 4x faster, and took the kiosk revamp live across 30+ outlets.
I walked into four channels, App, Web, Kiosk and WhatsApp, running without a shared product model. I rebuilt how we ship, and took the kiosk revamp end to end.
What it taught meClarity ships faster than headcount. The faster design turnaround didn't come from more designers. It came from deciding, in writing, how we make decisions.
Founded PS AnalytIQ, now in production, and cut counterfeit detection by 90%.
Counterfeiters were copying one genuine code onto thousands of fakes. I founded PS AnalytIQ to catch them proactively, and it's in production today. Alongside it, I built the company's inventory system from scratch and ran serialization for three of India's top five pharma makers.
What it taught meThe model was the easy half. Pharma is rigorous about documentation and procedure, and that discipline is what makes trust auditable.
Built energy models that told identical machines apart with 99% accuracy, and ran a state-wide water audit that became a peer-reviewed Springer paper.
Two missions. Real-time water-quality monitoring for the Jal Jeevan Mission, including an audit of Punjab's water we ran from a lab we set up ourselves. And e-Sense, NILM energy models built with German research partners.
What it taught meThink in systems, not components. Sensors, algorithms, hardware and people only work when you understand how each piece talks to the rest. And once you own the whole process, impossible timelines become logistics.
Went past writing test cases to improving the billing platform itself.
Quality automation for an analytics platform that bills a city's utilities. I didn't just document test cases, I improved the platform itself and kept its integrations honest.
My first exposure to ML that had to work in the field, not in a notebook. The start of everything above.
Here's what that looks like when it ships.
I rewrote how the team decides and ships, then took the kiosk revamp end to end: audit, redesign, prototype, ship.
I founded PS AnalytIQ, a hybrid ML engine with an in-house natural-language query layer, and took it from idea to production. It ships today as psAnalytiQ.
Real-time water-quality monitoring for the Jal Jeevan Mission, and e-Sense: NILM energy models sharp enough to tell identical machines apart, deployed with German research partners.
AI that ships inside a workflow beats a demo that impresses in a deck.
Machines take the manual work; people move up to the calls only they can make.
Knowing how AI works tells you where it will pay off, and where it won't.
The surest way I know to tell where AI pays off is to build with it myself. A few things I've been tinkering with off the clock, all public on GitHub.
A self-running weekly brief that ranks GitHub's trending repos by momentum, not lifetime stars, drops anything below a proven-star floor, then has an LLM explain each one in product terms: what it does, who it's for, and the week's theme. Built to answer a product question, not an engineering one.
deep_researchMulti-agent research · PythonGive it a topic and six small agents plan the searches, write a cited report, grade their own work, and re-research the weak spots before emailing the result. A compact, readable example of an agentic workflow with a self-correction loop and human-in-the-loop checkpoints.
rag-mfRAG pipeline · PythonA facts-only mutual-fund FAQ assistant, built in phases: it scrapes and normalises scheme pages, indexes them into Chroma with BGE embeddings, answers only from what it retrieves, refuses anything out of scope, and holds itself to a golden-set evaluation suite with daily re-indexing.
A peer-reviewed paper, and essays on AI and product.
The dataset behind it came from the Punjab groundwater audit I ran at CSIR-CSIO: hundreds of samples we collected and tested ourselves.
Read the paper ›Also writing on LinkedIn
We built minds we couldn't look inside. Anthropic's interpretability research just switched the light on, and what it shows is worth sitting with.
A Class 11 interview that rearranged how I think about AI guardrails. Users don't need protection, they need judgment.
Oh, and one more thing.
Silver at the Nationals as captain of my state team. Still playing: first runner-up at RHPL 2025.
Back at the school where I learned to skate. My students now win national medals of their own.
With the Harnoor Madhar Foundation: from COVID relief drives to getting 90+ village kids into school.
A distinction in Advanced Statistics on top of a CS degree. Still collecting subjects.
Or just . It knows where my calendar is.
Chats with my AI are logged so I can improve it. Please don't share personal details.
