KshitijJha

Why

I wanted to actually understand concurrency instead of trusting the runtime to sort it out for me.

What it is

Raw TCP, no framework. A listener hands each connection to its own goroutine, a shared register maps connections to users, and a mutex guards that register because several goroutines write to it at once. I got failures out of it: deadlocks, a concurrent map access race, WaitGroup misuse. I found each one by reasoning through the synchronisation model instead of adding sleeps until the crashing stopped. It's the smallest project on this list and the one that taught me the most.

Why

I wanted to know whether the strategies people post online survive contact with live data. Mostly they do not, and I wanted that answer from my own logs rather than somebody's screenshot.

What it is

Gold, silver and copper CFDs on the Capital.com API. EMA 9/21 crossover with RSI(14) confirmation, ATR(14) for stop sizing, a 1.5x ATR stop against a 3x ATR target, 1% account risk per trade, one open position per instrument. I fixed the stopping criteria in advance so I could not move them once it started losing. The lesson that stuck was infrastructural: GitHub Actions' built in cron was unreliable across a two hour test window, so the schedule now comes from an external trigger and the workflow listens.

Why

A university brief I refused to keep small.

What it is

A Tkinter GUI talking to a multithreaded TCP socket server over SQLite. The brief asked for none of that. I added live currency conversion across USD, GBP, INR and MUR, plus sales analytics in matplotlib. It was the first time I had two processes that needed to agree with each other, which is a specific and annoying kind of problem.

Why

An excuse to use XML and XSLT instead of treating them as legacy trivia, and to find out how messy real music metadata is. It is messy.

What it is

A record store front end where the album data gets scraped in Python from music APIs, structured as XML, then transformed into HTML with XSLT. The transformation layer does the job frameworks hide from you, which was the whole reason I built it that way.

Why

Hackathon, team of four, built against the clock. The idea was to push a policy through the perspectives of the people it lands on instead of the people who debate it.

What it is

A four stage pipeline sitting on top of an LLM. A classifier, then four specialist agents reading risk through a fiscal, labour, equity and regional lens, then up to 50 synthetic personas grounded in NSSO and Census 2011 demographics, then a coordinator that pulls the disagreements into one briefing. Most of the engineering is orchestration and prompt design rather than novel modelling. The hard part was batching the persona calls so the thing returned inside a demo slot. We sanity checked it against MGNREGA, which has twenty years of published outcomes to check against.

Why

An experiment in whether a prediction market prices weather badly, and an excuse to read about forecast verification, which I knew nothing about going in.

What it is

It blends ECMWF, GFS and UKMO ensembles with two deterministic providers into a single distribution, adds a climatological prior weighted by how far out the forecast is, and only logs a prediction when its edge over the market clears ten points. Everything resolves against historical records and scores itself on win rate, ROI and calibration. It is still running. The honest verdict is that the sample is nowhere near big enough to claim an edge.

Why

Campus transport was expensive and students were already driving the same routes anyway. The hard part was trust, not the software.

What it is

A ride sharing platform pairing verified student drivers with riders. Driver verification, a dashboard with listings and analytics, route filtering, ratings, all on Firebase Auth, Firestore and Storage. Built with Harshil Patel and shipped to a live campus audience.