DRDaniel Roberts
Chimera · autonomous markets platform

Chimera, an autonomous, self-auditing trading research platform

I built Chimera, an autonomous, self-auditing multi-asset research and execution platform. It runs a self-healing service fleet unattended, holds a 5+ TB proprietary corpus, and proves every strategy against real execution costs, out-of-sample, before capital touches it. Live where the gate has passed. Capital deployment is decided per strategy, not per platform: two strategies are live today (futures on Coinbase, event contracts on Kalshi), with the wider options bench (tastytrade) in gated paper rollout. Every strategy earns promotion independently through the same pipeline (gate → shadow → canary → live) and gets demoted the same day it fails re-validation. Strategies rotate in and out daily as evidence changes; the corpus grows on schedule as collectors come online in their windows. Paper isn't a caveat here; it's a stage every strategy passes through, and returns to if it stops earning its place.

158self-healing services
5+ TBresearch corpus
4venues instrumented
88k+configs gated
Universe

One stack, every asset class.

Roughly 200 optionable underlyings inside a ~490-name daily research universe: crypto, equities, ETFs, indexes, index and energy futures, metals, and FX. The same name trades on several cadences at once: SPY runs in the daily fleet, the 0DTE intraday fleet, and the options IV-native fleet, each with its own models, horizon, and validation track.

F1Crypto 5-minute
BTC, ETH, SOL perps and spot on Coinbase, Kraken, and Binance feeds; the lowest-latency book. Best of: GRU SOL vol-spike 0.828 · TFT ETH 0.801 · JEPA BTC 0.805 · JEPA owns crypto crash on all three.
F2Daily universe
~490 equities, ETFs, indexes, futures, and FX pairs with decades of history per name. Best of: N-BEATS HYG vol-spike 0.906, the best non-proxy cell in the fleet · TabPFN ^NDX 0.899 · volatility and magnitude edges run 0.65 to 0.90 across the board.
F30DTE intraday
~190 liquid options underlyings: SPY, QQQ, IWM and the index complex, same-day expiry structures. Best of: inside-move 0.911 · spread-safe 0.884 · gamma-state 0.986 (flagged proxy, disclosed).
F4Options IV-native
~200 underlyings with full Greeks and IV-surface history from the tastytrade chain archive. Best of: IV crush 0.907 (TabPFN) · IV expansion 0.860 (Mamba) · premium survival 0.801 (XGB) · CatBoost’s best IV-crush cell 0.888.

The strategy bench is organized as eleven sleeves, each named for the Greek letter of the risk it monetizes, weighted to options and futures, not crypto: vol-expansion long premium (straddles and strangles across weekly, monthly, and 0DTE tenors), defined-risk short premium, IV term-structure calendars, 0DTE pin & delta-decay structures, second-order Greek candidates, rates & macro event contracts, and directional index-futures candidates, every one walking the same gate → shadow → canary → live path. Two sleeves are live (alpha, beta), five more are in gated paper shadow (charm, delta, gamma, theta, vega), four remain in gate, and demotion lands the day re-validation fails. The live page publishes each sleeve’s evidence, including the failures.

System map

Five layers, strictly ordered.

L1Data spine
20+ collectors: Coinbase L1/L2 depth, Kraken, Binance spot+derivatives (WS-persisted), Kalshi perp order books, funding, open interest, liquidations, ETH/SOL on-chain, news firehose, sentiment. Stores: TimescaleDB hypertables (unified 1-second features, raw trades, order-book deltas, 1-second silver layer) · Redis hot state · daily parquet lake.
L2Feature & ML plane
GPU feature computer (LPPL, regime, vol zones, VPIN, Hive consensus) · vector assembler (1-second fast features) · a twelve-architecture forecast fleet scored on honest walk-forward AUC, where the TabPFN family holds the plurality of championships (804 of 2,717 rows, 2.0× the next family) and a majority of podiums (53.7%) · ZoneX: 10 per-volatility-zone exit networks · XGBoost sleeves · regime classifier · leakage-guard watchdogs.
L3Research factory (the core)
Numba backtester with calibrated real costs · GPU sweep infrastructure (80k-cell searches in minutes) · scientific gate: purged+embargoed walk-forward, deflated Sharpe, three null controls, both-halves, stress perturbation, feature-basis robustness · forward book with verdict rules frozen before first data.
L4Execution
Multi-clock hot path: GPU decision path <100µs · fast decision path <50µs · order dispatch 100–250 ms venue round-trip · shadow runner (live+paper) · maker/taker routing, order splitting, retry & fill reconciliation · venue clients: Coinbase CFM, Kraken futures, Bitnomial, Kalshi. Live: Coinbase futures + Kalshi event contracts; tastytrade options in gated rollout. Risk stack: capital guardian + DEFCON stress gates + drawdown throttle.
L5Ops & self-healing
Infra-heal (auto-restarts critical containers, ~38 s recovery) · heartbeat watchdogs across the fleet (24/26 fresh at the last published snapshot) · table-freshness & parity watchdogs · nightly ops loop · live monitor with true-forward vs backtest-seed columns per experiment.

Full architecture, evidence record, and honest economics in the white paper. Operational specifics redacted for the public edition.