Projects
Autonomous trading system for Polymarket's hourly crypto markets
Problem.
Polymarket’s hourly BTC/ETH/SOL markets resolve on a known reference price every hour. Most participants trade them on gut. I wanted to see whether a structured, intelligence driven system could hold an edge over hundreds of trades when provided realmarkets data.
What I built.
- A bot that pulls hourly market odds, queries a panel of specialist AI prompts trend regime momentum through Brain API, and turns their consensus into a confidence score.
- Deterministic gates on top: ADX based regime filter, skip rules, position sizing. The model advises code decides whether and how much.
- Every decision entry, skip reason, confidence, outcome logged to Notion so the whole history is auditable.
Previously, this experiment achieved a total gain of 200% over deposited capital. Check more details. After the introduction of trading fees on Polymarket, the strategy was no longer profitable and has therefore been paused.
Fleet monitoring and anti-theft platform for mining contractors
Problem.
Small mining contractors in India (10–50 vehicles) lose money to diesel theft, false trip reporting, and inflated work-hour claims, and the only fix available today is the owner physically standing on site. Generic telematics tracks GPS but has no mining workflow; enterprise mining fleet systems solve it but are priced far out of reach for a small contractor.
What I built.
- Custom ESP32-S3 GPS trackers with store-and-forward buffering (flash ring buffer, replay-on-reconnect) speaking the standard OsmAnd protocol into a self-hosted Traccar server.
- A FastAPI backend holding all the proprietary logic: a per-vehicle trip state machine that turns raw GPS into verified load→haul→dump cycles, dynamic excavator “fill zones” for automatic load attribution, and an events/alerts engine (night movement, boundary exits, overspeed, breakdowns) all built anti-flap/anti-jitter from the ground up after real false positives surfaced during testing.
- A simulator that emulates a full fleet over the real device protocol against a real backend, used as the regression suite: it has already caught and fixed real concurrency bugs and detector edge cases (GPS jitter at zone boundaries, dropouts spanning trip transitions) before any of it touched actual hardware.
- A React + Leaflet dashboard (live map, vehicle detail with route replay, zone editor, live events feed).