Go to the place.
Then measure.

I went out to sit with Farmer Producer Organisations expecting to find a data problem. There wasn't one. Soil data existed. Crop guidance existed. What didn't exist was a person with the time to translate one into the other — the extension worker who was supposed to visit and never could, because there is roughly one of them for thousands of farms.

So Vidhai isn't a recommendation engine. It's an attempt to be that missing person. Parallel agents search a live datastore across ten semantic paths, score candidate crops on soil health and the economics of switching, and — the part that mattered most in the field — let a farmer argue back. Swap a crop mid-conversation and the whole plan regenerates around the change, because a plan you can't push back on is a leaflet, not advice.

Built on Google ADK with Gemini 2.5 Pro, containerised onto Cloud Run. It won the Google Cloud × University of Birmingham hackathon, which was nice, but the thing I keep is what the fieldwork taught me: the bottleneck is almost never where the dataset is.

Ladakh loses its water to timing. The snowmelt arrives after the planting season needs it. Ice stupas are the local answer — artificial glaciers, sprayed through winter, that melt exactly when spring asks. I spent a winter under Sonam Wangchuk's institute trying to make them predictable rather than hopeful.

The interesting engineering wasn't the model, it was the place. Raspberry Pi and ESP-32 nodes on a glacier, with intermittent power and no inbound network, streaming telemetry out through Cloudflare Tunnels because there was no other way through. Then a multivariate autoencoder and isolation forest replacing the rule-based thresholds that kept firing on weather rather than on faults, with LSTMs forecasting the freezing window.

Everything I now believe about deployment I learned from a machine I could not SSH into at −20 °C.

A summer spent below the abstraction layer: writing a GNSS software-defined radio from scratch, implementing GPS L1 C/A acquisition and tracking against raw intermediate-frequency sample streams. No framework to hide behind — correlators, Doppler bins, and the slow realisation that a wrong answer arrives just as confidently as a right one.

I cut processing latency 35% by vectorising the matrix work through NumPy. But the lasting part was learning to distrust output that looks plausible, which is the entire subject of the next chapter.

An engineer asks a document system for a tolerance and gets a number with a citation attached. The citation resolves. The number is wrong. Nothing in the interface distinguishes that from the truth — and in manufacturing, that costs more than an unanswered question.

So I built a copilot that treats a citation as a claim to be checked, not decoration: layout-aware parsing, ColBERT late-interaction retrieval, a local 4B model, and a verification gate — M_eng — that every answer must pass before a human sees it. On the published ALCE benchmark it recovers +27.7 citation recall over a naive-RAG baseline. An answer that fails the gate is withheld, not shown with a warning.

Try it. One of these three sentences is not supported by the source it cites.

That third sentence is the whole problem in miniature. It is fluent, it carries a citation that resolves to a real page, and it is unsupported. Word-overlap scoring waves it through, because every word in it appears somewhere in the source.

Which raised the obvious question about my own metric. So I had an independent model, blind to the score, adjudicate every answer the gate had ever approved on an engineering document — all fifty-three of them, a census rather than a sample, with nothing left for me to have chosen.

Four in five held. The interesting number is the third one. I expected the failures to be marginal — answers that scraped past a threshold I could tighten. Every single one scored a perfect hundred, identical to the answers that were right. There is no number anywhere in the system that separates them.

That is a more useful thing to know than the 92% I could have led with. A metric that is noisy can be recalibrated; one that is blind has to be replaced. And it is blind to a specific thing: of the answers that got through, only two invented a number at all. The rest were real values at the wrong address — right figure, wrong page, wrong document. Scaling the corpus from one manual to ten made it plain: thirty percent then cited the wrong manual entirely, a failure the single-document pilot could not express. I would rather ship a system whose limits I can name.

Also: 63s → 30s local latency with byte-identical outputs, and a hosted build at 100 MB resident, down from 6.1 GB. Live demo · Source

I'm finishing the MSc in September 2026 and looking for founding-engineer or forward-deployed work where the problem is still ambiguous — where someone has to go find out what's actually wrong before anything gets built, and then be honest about whether it worked.

Distinction (76.5%), top marks in Neural Computation (84) and NLP (83). Chancellor's Scholarship, one of ten awarded globally. Published in MIND 2024, Springer Nature, on real-time Tamil fingerspelling translation. Previously coordinated logistics for 100,000+ contacts at the Isha Foundation, which is where I learned that systems are mostly people.