Trading Platform

A 24/7 automated trading platform I built to run a systematic, model-driven strategy. It generates ML-based entry signals, routes orders through a FastAPI execution engine in under 200 ms, maintains a full SQLite ledger with idempotency keys, and exposes a live React dashboard for monitoring positions, P&L, and system health.

Motivation

I wanted hands-on experience with the engineering problems behind real trading infrastructure. Not just the ML side, but the execution layer: latency, fault tolerance, partial fills, and audit trails. Building it myself forced me to solve problems that are glossed over in tutorials.

Execution engine

The core is a FastAPI service that processes order intents from the signal layer, validates them against position limits and available balance, and submits them to the exchange. End-to-end order latency is consistently under 200 ms. Idempotency keys prevent duplicate orders if a request is retried after a network failure. Partial fills are tracked in SQLite and reconciled against expected fills before the system places the next order.

Signal generation

The signal layer runs a set of Python models trained on price and volume features. Models are evaluated periodically and new ones are promoted to live only when they pass backtesting thresholds. The architecture keeps signal generation and execution decoupled: a signal is just a JSON payload with a direction and confidence score, so the execution engine doesn't know or care what produced it.

Infrastructure

The platform runs on a VPS behind a WireGuard VPN. No open ports except the VPN tunnel. The React dashboard connects through the same tunnel, so position data and system health are visible from anywhere without exposing the execution engine. Discord alerts fire on fills, errors, and daily P&L summaries.