Nine signals. Five agents. One trading brain.
Most platforms ship one model and pray. We ship nine — each independent, each shadow-validated for 28+ days before it touches your capital. The orchestrator blends them, the kill-switch overrides them, and an LLM explains every trade in plain English.
Every signal earns its weight — or runs in shadow until it does.
New signals ship at weight = 0. They have to beat the locked Phase 37 baseline for four consecutive walk-forward weeks before they get a real allocation. No exceptions — including for our own models.
The locked walk-forward validated baseline. Carry 10% / trend 90% sleeve-class quota, James-Stein shrinkage, ATR-targeted vol overlay. Sharpe 2.39, MaxDD -10.51% OOS.
HMM over BTC + cross-asset features. 4 regimes (trend / chop / risk-off / quiet) gate every sleeve. Disagreement automatically scales position size down.
Funding-spike breaker + DD-brake per sleeve class (trend 15/5, carry 8/3, xcarry 6/2). Halts entries before the regime change finishes.
Champion / challenger policies trained nightly in our Python research lab on Fly.io. Outputs a signed score [-2, +2]. Shadow weight today — flips on at 28d green.
Deribit BTC/ETH chain snapshot every 15min. 25Δ skew + ATM IV term structure feed a directional signal. Data-only today.
5 free RSS feeds → text-embedding-3-small → Pinecone market-news vectors. Per-symbol bull/bear cosine gates entries.
5 specialised agents (Trend / Mean-Rev / Vol / Macro / Sentiment) debate every 15min on BTC/ETH/SOL/BNB/XRP. Trade fires only when agreement crosses threshold.
Gradient-boosted classifier over every closed trade. Outputs Platt-calibrated P(win ≥0% over 5d). Retrains weekly, shadow today.
Regime-conditional sleeve performance + multi-timeframe BTC composite + real-time PnL attribution. Shipped 2026-05-20.
Five AIs that disagree — so you don't lose.
The Multi-Agent Swarm runs every 15 minutes on BTC, ETH, SOL, BNB and XRP. Five specialised AI agents independently analyse the same market and vote long / short / neutral. The trade only fires when agreement crosses threshold. Single-model failure mode — eliminated.
Unlock the Swarm — Alpha planMomentum, breakouts, Donchian / ATR — the disciplined trend-follower.
Bollinger, RSI extremes, overbought/oversold — the contrarian.
IV term structure, realised vol, regime classification — the vol trader.
DXY / US10Y / Gold / VIX cross-asset — the macro analyst.
News flow, social, Pinecone narrative vectors — the tape reader.
Strategy Evolution — the engine that writes its own edge.
Most quant shops have one team writing strategies and one CIO deciding which survive. We replaced both with a genetic algorithm that breeds, tests, and retires strategies on a continuous loop — and pipes the survivors through the same statistical-rigour gates a hedge fund applies to a new manager.
Every cycle, ~20 new candidates are spawned from crossover + mutation of the top-quartile survivors. Population diversity is enforced — no strategy monoculture.
Each candidate must pass DSR (Deflated Sharpe), White's Reality Check, and PBO (Probability of Backtest Overfitting) — the same tests CIOs use to evaluate external managers.
Qualified → Paper → Canary → Live. Auto-kill on Sharpe degradation or DD breach. Phase 37 baseline stays locked — Evolution is additive.
Live trading is unchanged. Evolution candidates sit in shadow with weight 0 until a human admin explicitly promotes them after 28+ days of OOS validation.
- Continuous adaptation — when regimes shift, new genomes mutate to match before human researchers notice.
- Population-level robustness — single overfit candidates die at the stat-gate. Only edges that survive multi-test scrutiny graduate.
- Honest in-distribution validation — White's Reality Check runs against the FULL search universe, not the survivor-bias subset.
- Full audit trail — every promotion, kill, and parameter change writes to immutable lifecycle transitions.
- Composable with the orchestrator — surviving strategies enter the 9-signal blend as a 10th conditional signal at weight 0 until they earn it.
Six layers between a signal and a fill.
Platt-calibrated ensemble over 9 signals. Selective AI-overlay size multiplier 0.5–1.0×. Three-flag gate (wf_v2_gate_open + per-user opt-in + global kill-switch).
Cold-start cap → proven → scaled ladder. Per-user circuit breaker. Order anomaly detector. Global 20% DD breaker, fail-CLOSED on every check.
Smart Order Router across 13 exchanges, adaptive TWAP from learned slippage, maker-first with adaptive wait, DRL slicing agent.
Every trade gets an LLM-generated rationale in plain English. Cron-batched 15min for backfill. Searchable via /admin/decision-shadow.
Prometheus exporter + Grafana Cloud Mimir push + ClickHouse traces & logs. 11 SLOs with PagerDuty + on-call rotation.
Cloudflare Workers + Durable Objects + Queues. Rust→WASM indicator kernels. 3-tier price cache (L1 isolate / L2 compute_cache / L3 origin). NATS event bus.
No one goes from zero to $100k overnight.
Live trading walks a defined ladder. Each tier requires real proof — Sharpe, drawdown bounds, days in market — before the next size unlocks. Two-person approval at the top. The runbook is public.
Cold-start. 14 days minimum. Position cap 5% per name.
Proven — Sharpe ≥ 1 over 14d, MaxDD within bounds.
Scaled — Sharpe ≥ 1.5 over 30d, cumulative PnL > 0.
Institutional. Two-person approval + 24h cooling-off.
Phase 37 walk-forward hypothetical backtest, locked baseline. Re-validated weekly. Not live trading.
Worst peak-to-trough in 12 months OOS. Hard 20% global breaker above this.
Hypothetical Phase 37 walk-forward backtest, last 12 months @ 10 bps slippage. Not live trading.
An edge that compounds while you sleep.
- Non-custodial — your funds never leave your exchange
- 13 exchanges supported (Binance, Bybit, OKX, Coinbase, Hyperliquid, …)
- Withdrawal-permission audit auto-revokes any key with withdraw enabled
- Two-person kill-switch — no single admin can restart live trading alone
- Every trade logged with a human-readable LLM explanation
- Walk-forward validated baseline you can re-run yourself on /engine
Stop guessing. Start compounding.
7-day Pro trial. No card required. Cancel anytime. The AI keeps running whether you watch or not.