An AI trading bot,
pipeline-first.
Here's how the stack actually trades: a cross-asset regime call, a watchlist built for that cycle, a fat per-symbol snapshot (bars, enrichers, data-quality nagging), a small model pitching a JSON-shaped move, a risk engine that can veto or soften it, an executor that talks to the broker, and a journal that remembers every bruise in SQLite + JSONL. It's wired for paper first as a science project and a spectator sport, and I'm genuinely curious how much hypothetical money it can set on fire over a long enough timeline if we keep the lights on.
Pipeline flow
Pan, zoom, and click any node in a Mermaid diagram of the per-cycle pipeline: external feeds, regime, snapshot, brain, risk, executor, and sinks.
Open the diagram
Architecture prose
Sanitized walkthrough: runtime topology, market snapshot, model decision flow, risk engine policies, executor guards, journal layout, weekly reporting.
Read the doc
What's inside
- Cross-asset regime pre-brain (TTL-cached cross-asset verdict)
- Watchlist construction: screener + anchors + bearish hedges + held positions
- Snapshot builder with data-quality and sector context
- Tier-1 / Tier-2 risk policies and intraday giveback halts
- Executor: bracketed equity orders + single-leg options + liquidity gate
- Journal: SQLite + JSONL with size-based rotation and WAL concurrency
- Weekly Sonnet review → human blog + trader feedback bullets