WealthPilot

When Ashira Capital Partners approached us, they had a clear problem: their analysts were drowning in signal noise. Congressional trades, dark pool activity, earnings calls, Reddit chatter, SEC filings — the data sources that move markets were too numerous and too fast for manual synthesis. They wanted a system that could ingest it all, make paper trades autonomously, and — critically — improve itself over time without human intervention.

We built WealthPilot in Python and deployed it on a VPS managed by PM2, running hourly trading cycles from market open to close (9:30-16:00 ET) via APScheduler. But the architecture underneath that simple schedule is anything but simple. At the core of WealthPilot are 10 proprietary edge tools, each tracking a distinct alpha signal. Committee Alpha follows congressional trading disclosures. An earnings hallucination checker uses Claude's API to cross-reference CEO claims on calls against actual financials — flagging discrepancies before the market prices them in. An options whale-watcher and dark pool monitor surface institutional positioning. Reddit social sentiment, LLM-driven news analysis, SEC Form 4 insider filings, short interest data, government contract awards, and analyst rating changes round out the picture. These signals feed into a scoring engine built on 5 technical indicators — RSI, EMA crossovers, MACD, volume profile, SMA, and ATR — producing a composite conviction score for every candidate trade.

We layered 10 distinct risk controls. Kelly criterion sizing prevents overallocation. ATR-based stops and trailing stops manage downside dynamically. Partial exits lock in gains incrementally. A circuit breaker halts trading entirely if drawdown thresholds are hit. Regime-adjusted position sizing shrinks exposure in volatile environments. Correlation penalties prevent concentration risk. Earnings blackout windows keep us out of binary events. What separates WealthPilot from a static rules engine is its feedback architecture. Every trade is logged, and signal weights adapt based on recent performance. A daily discovery scan surfaces new tickers meeting edge criteria. An intelligence scanner re-evaluates the macro regime. Weekly evolution cycles retrain the ML filter — a logistic regression model trained on 436 samples with time-decay weighting, validated through walk-forward out-of-sample testing. Monte Carlo simulations stress-test the portfolio against historical shocks: COVID, the 2022 rate cycle, the SVB collapse. An autonomous monitoring agent we call Plutus oversees the system around the clock, with full kill switch authority.

WealthPilot ran live paper trading through Alpaca on a portfolio starting at $100K. We wrote 497 tests across 32 test files — not because we love testing, but because an autonomous system that trades money while you sleep demands that level of coverage.