We start with raw data — ~70 vendor and alt-data sources continuously ingested
into Postgres + Redis + QuestDB. Polygon for OHLCV, SEC EDGAR for filings, Quiver and
House Stock Watcher for congressional trades, 13F amendments for institutional flow,
NIM for news, Reddit + StockTwits for sentiment, GPW + ESPI for the Polish market.
On top of that runs an 8-model ML ensemble — chart-CV v3 (3-branch
fusion: ViT + PatchTST + Tabular), LSTM with attention, an earnings-surprise model,
Chronos T5 zero-shot probabilistic forecasts, a 4-regime classifier, FinBERT sentiment,
a smart-money composite, and an ensemble that combines them. Every prediction is
settled publicly on our calibration page.
The output is fed into a decision engine that runs every 15 minutes, talks to
multi-provider LLMs (Claude Opus 4.7 primary, kie.ai fallback), parses strict JSON with
Zod, enforces risk gates (size cap, stop/TP required, drawdown hard-stop), and writes
actions to a paper trader or to real brokers (MEXC, Interactive Brokers, CCXT) with
auto-failover and per-symbol routing.
Layered on top: a verified-P&L community, a Pine-like DSL (TraderPeak Script), the
full Polish-tax stack (IKE / IKZE / OKI cascade, PIT-38 multi-broker importer), and a
backtest robustness panel that runs CPCV, Bailey/LdP Deflated Sharpe, PBO and Monte
Carlo bootstrap by default — because we'd rather flag a fragile backtest than
show you a sexy equity curve.
~70
Data sources
Vendor + alt-data feeds continuously ingested.
8
ML models in production
All deployed on RunPod, all settled publicly.
14,000+
Tickers tracked
US equities, ETFs, crypto, forex, options, GPW.
78
DB migrations
Built over 18 months — not a wrapper around someone else's API.
