Articles
Deep dives into how the system actually works — written by the developer, grounded in the code, corrected in public when the code changes.
Latest
How the bot reads charts: from candles to conviction
The pipeline from raw market data to a reasoned decision — market history, a 50-level order book, Numba-compiled indicators, a 1920×1080 chart image, and deterministic risk limits before an order is sent.
Vector memory: how an LLM trader remembers its mistakes
Closed trades become 768-dimensional vectors in ChromaDB. Before each decision the bot reviews its own history — 3 nearest setups, 20-trade statistics, 5 recent rejections — and a math check separates luck from skill.
From paper to a live executor: why order placement is a separate service
Reasoning and order execution are deliberately decoupled. The seven checks every decision passes, the verdict journal that closes the loop — and the honest state of the live path: testnet only, reconciliation still the weak link.
The full development story (December 2025 – September 2026)
The engineering history with real commit dates, the mathematics behind the gates, an honest status report, and a list of the claims this site got wrong and had to fix.
See the system in action: the live dashboard streams every decision, prompt and reasoning step in real time.
Live dashboard ↗GitHub ↗