agents, ios apps, infra, SaaS, backtester and finance.

every diabetes app assumes you will weigh your food and look up each ingredient. she was never going to do that, and neither would i.

so it takes a photograph. one pass identifies the dishes, a second estimates weights and nutrition ranges. ranges rather than fake precision, because a photograph cannot tell you grams.

everything stays on the device. sync is optional and off by default. a health record for one person did not need a backend.

google photos is genuinely excellent and the price is handing over every photograph you have ever taken.

this runs the same class of analysis locally, so the library becomes searchable by what is in it rather than only by when it was taken.

it is slower than the cloud version and always will be. that is the trade, stated plainly.

three apps, two remotes, and a four second cold start to switch off a light or drop the ac two degrees. the hardware was fine. the software between me and the hardware was the problem.

wiz bulbs speak a simple protocol on the local network, so talking to them directly skips the round trip to a cloud service entirely. the ir blaster covers the ac and everything else older than wifi, so one interface reaches the whole room.

one screen, one tap, no account. the smallest useful thing i have built, and the one i use most.

i have been the unofficial family portfolio manager for a while, which meant screenshots and p&l summaries on request, all day, forever.

this fetches all of them and shows one position list with live prices. no advice, no charts, no engagement mechanics.

it is deliberately boring. that is why it still gets opened.

the college portal tells you your attendance percentage but not the thing everyone actually wants to know, which is how many more classes you can afford to miss.

so it computes that directly, per subject and combined, and lets you correct a wrong record with a long press instead of an email to an administrator.

nothing leaves the phone. an attendance record is not worth a server, an account system, or a privacy policy, and a student should not have to trust one.

after any large event there is a shared drive with two thousand photographs in it, and finding the eleven you appear in means scrolling all two thousand.

one scan returns only yours. the recognition is the easy part now; the useful part is framing it around a person rather than an album.

built after a wedding, for the obvious reason.

you have to identify a component before you can look anything up, and the identifying marks are often worn away or facing the wrong direction.

yolov8 handles recognition, llama 3.1 handles the conversation, and the answer is grounded in bharat electronics own manuals rather than invented.

the constraint that shaped it: a confident wrong answer about a defence component is far worse than admitting uncertainty.

every night a metro depot decides where each train parks and in what order they leave. get it wrong and you are shunting trains at four in the morning.

certificates and maintenance windows constrain what can run at all. branding contracts want particular trains on particular lines. mileage wants to stay even so nothing wears out first.

real constraints and no clean optimum, which is exactly why it was worth doing.

uploading an entire camera roll to find the receipts in it is expensive, slow, and a privacy problem you then have to explain to a client.

so a mobilenetv3 classifier runs on the device and filters first. only likely receipts are sent up, which cuts the bill and means holiday photos never leave the phone.

behind it sits the unglamorous half: idempotent uploads, job claims, backoff retries, background reconciliation, and observability with the personal data scrubbed out. that is what makes it survive real users.

a portfolio is a strange brief. the work is the point, so anything the site does to draw attention to itself is working against the person paying for it.

so this is deliberately quiet. fast to load, easy to read, and structured so the work is the first thing you see rather than an intro animation.

built in next.js with the motion kept to almost nothing. the only thing it needs to do well is get out of the way.

this one was not a portfolio in the usual sense. the brief was a brand, so a list of projects would have been the wrong shape entirely.

the site is built around identity first: the tone, the palette and the way it reads matter more here than any feature.

kept simple on purpose, because a brand site that needs explaining has already failed.

backtesting at tick resolution is where a personal project usually dies. the naive version takes hours per run, so you stop iterating, so you stop learning anything.

rust turned it into something i could run between edits. the point was never a profitable strategy.

scoped to one instrument on purpose. gold against the dollar is volatile enough to be interesting and liquid enough that the fills are not fiction.

indian credit underwriting still runs on an analyst reading a pdf and typing ratios into a spreadsheet. it is slow, inconsistent, and the reasoning evaporates the moment it is finished.

this pulls around 25 financial fields from any annual report, computes tier 1a ratios, then runs two models trained on a cohort of wilful defaulters against matched controls.

the output is a memo, not a score. a number nobody can interrogate is worse than no number.

reading the alphafold paper and understanding it turned out to be different activities, and the gap between them was code.

so this rebuilds the data preparation pipeline at a scale that fits on one machine, with a dashboard comparing its toy inferences against real traces.

it does not fold proteins competitively and never intended to. it answers a narrower question.

most agent frameworks fix the strategy at design time. when it fails it fails the same way forever, and you are the one who has to notice.

this keeps a lightweight reinforcement learning bandit over its own approaches, so tactics that keep working get chosen more often.

a small idea tested honestly rather than a framework. the interesting result was how fast it learned to stop doing the thing that never worked.