David Kaye

End-to-end automated trading platform targeting 1–60 minute horizons on liquid crypto perps (Binance Futures, Bybit, KuCoin). Six typed, contract-bounded components — Feed, Feat, Book, Eval, Exec, and a shared Core — compose into a single hot-path trading process plus an offline evaluation pipeline that re-runs the exact same production code over Parquet replays. Designed against explicit risk targets (Sharpe ≥ 2.0, <5% monthly max drawdown) with a gated go-live process: backtest → walk-forward → stress → paper → capital ramp.

Built in Rust for deterministic latency; Python (PyO3) is used for Bayesian hyperparameter search via Optuna and GBM meta-model training (XGBoost / LightGBM). Deployed on Kubernetes, with multi-sleeve architecture (CEX first; DEX/options/news-driven sleeves on the roadmap), market-partitioned trading processes, and a roadmap that layers in cross-asset macro leads, options-derived skew signals, funding-carry, and LLM-based news/sentiment agents.

A new code-generation backend for the Idris2 dependently-typed language that replaces the stock Chez Scheme runtime with native ML-family targets, producing standalone binaries that run 7–20× faster than Idris2's default output across the benchmark suite.

TLC intercepts the pipeline before type information is discarded: it walks the typed Term IR, drives Idris2's own partial evaluator to a fixpoint to monomorphise every polymorphic call site, and reifies the result through a new typed lambda-calculus IR (TExp) that targets Haskell, F#, and OCaml — all via Idris2's existing Codegen extension point, with zero forks of the upstream compiler.

Across nine benchmarks (fib, par-fib, tree-depth, insertion-sort, line-count, list-sum, multi-mod, file-scan, higher-order), OCaml-emitted binaries finish in 2–51 ms versus 50–411 ms on Idris2's Chez runtime.