Four local personas + one cloud host, paper-trading on Alpaca, transparent reasoning, public outcomes.
The MoE ensemble runs entirely on owned hardware. No OpenAI bill, no Anthropic bill, no per-token costs. Every debate is free at the marginal compute level, making daily content economical.
256 GB unified memory. Runs the 4-persona ensemble locally. Hosts the FastAPI backend, dashboard, and Alpaca integration.
128GB unified, 1 PFLOP FP4. Generates show assets, Flux 2 portraits, Wan 2.2 video, for episode production. Also runs an optional second-opinion judge for high-conviction calls.
Commission-free broker, JSON API. Every order is paper-traded. Operating cash flow funds the show, not viewer money.
Real ML architecture, not a marketing word. Four locally-run personas with different priors argue every stock pick. Bull team vs Bear team. Asymmetric verdict logic decides the trade. Donald (cloud) grades the panel after each cycle.
80+ free data sources including SEC EDGAR, FOMC, news sentiment, options flow, congressional trades, court records, insider transactions, app-store reviews, Federal Register, USAspending, developer activity, Telegram signal channels and more, full list at /data.
Verdict + confidence + slot + already-held + sector cap + drawdown + wash sale + correlation + event blackout + statistical filter + multi-timeframe.
Stops sized to volatility, not fixed percentages. High-vol names get wider stops, low-vol names get tighter.
Every closed trade retroactively grades the originating verdict. Per-persona walk-forward weighting. Monthly prompt evolution.
Every verdict is recorded. Every trade is visible. Every loss is published. No cherry-picking, no marketing-massaged numbers.
You're watching real reasoning produce real outcomes, not a presenter explaining what happened. The AIs do the work and the talking.
Survivorship-corrected backtests on 25 years of S&P 500 history. Modest, defensible numbers, published openly when validated live. Not get-rich-quick. See methodology →
The bot's code, the data pipeline, eventually the trained model, all open source. Free to fork, study, run yourself.