Systems that are actually running. Some is client work, so there is no repository to show. Where the code is mine and public, it is linked.
$200M private credit fund // Vantedge AI
One model does the extraction, a second model checks it, Python handles every calculation, and each figure links back to the exact spot on the scanned page it came from. Built for a fund where a wrong number is not a rounding error, and where the output had to reconcile against an institutional administrator.
100% on the benchmark set, around 99% in production.
Personal build
A voice agent built end to end. OpenAI Realtime for the conversational loop, Silero VAD, with barge-in and adaptive echo cancellation tuned for an actual phone leg rather than a clean 24 kHz demo.
The hard part was never the model. It was working out when someone genuinely wants to interrupt, and when they have just paused to think.
$200M private credit fund // Vantedge AI
Reads more than 100 badly scanned court documents per case and fills out the due diligence document used to underwrite it. Scans that come back with no text at all are where these systems quietly fail, so most of the work lived in the intake and verification layers rather than the model.
99+% accuracy.
Personal agent infrastructure // open source
An orchestration layer I run my own work on. A master agent spawns and directs sub-agents through a local broker API, wired into WhatsApp, email, a browser and a scheduler. It handles the repetitive half of my job.
It also applied to several of the jobs that led to this page, which is either a good sign or a worrying one.
Built 2023, self-taught
Scrapes filings and announcements off the National Stock Exchange and pushes them to Telegram in about two minutes, against an hour or more for traditional media to report the same thing. The first real thing I built after teaching myself Python.
It made a few hundred dollars, which at the time felt enormous.
Still running, on two channels: NSE Live News and NSE Financial Results Live.
In progress
A real-time voice agent that teaches accounting by asking rather than telling, then generates and marks an end-of-lesson test through a live function call. Built on the same voice stack as the sales agent, re-prompted and re-scoped for teaching.
Four years of teaching accountancy, pointed at the thing I now build.
Built 2023
Scrapes and standardises equity research from Indian brokers, then backtests each analyst's past calls to score how accurate they actually turned out to be, producing a smart money sentiment signal. Retail investors in India cannot see street numbers for most stocks.
Open source
Document retrieval using full-document density heatmaps over FAISS indexes. Instead of blind top-k search, it scores every page against several seed terms first, so the model picks pages by spotting clusters rather than guessing.
Open source
An agentic RAG research assistant over Indian annual reports.