AI Trading Bot

The Question

Can a disciplined system outperform emotional trading?

After years of watching myself — and every trader I know — make the same mistakes at the same points in market cycles, I wanted to build something that trades on rules, not feelings. Not because I believe markets are efficient, but because I know I'm not.

What I Built

A Python-based algorithmic trading system that runs against Bybit in paper mode. 50,000 lines of code, 266 passing tests, fully Dockerised with CI/CD.

The architecture is built around three principles:

• Signal generation — multiple data sources feeding into a thesis-aware scoring system
• Risk management — position sizing, stop losses, and portfolio-level exposure limits that cannot be overridden
• Execution — automated order management with slippage protection

The hardest lesson wasn't technical. It was accepting that the bot is only as good as the thesis it's built on. No amount of engineering sophistication compensates for a wrong read on market structure.

What I Learned

Building a trading bot is building a mirror. Every bias, every shortcut, every assumption you've internalised about how markets work — you have to make it explicit and write it down as a rule. And when the rule doesn't work, you can't blame emotion. You have to confront the gap between what you think you know and what the market actually does.

Status

Paper-only, and paused. The research loop and the paper trader run on a VPS; no live capital has ever been committed. That is a deliberate gate rather than an unfinished feature — the rule is fourteen consecutive days of clean paper results before live is even considered, and --mode live falls back to paper until it clears. Active development stopped in June 2026.

I'd rather say that plainly than let a "live" badge imply money is at risk when it isn't.