Work

Took 2,600 production pipelines from 8% to ~60% memory utilisation by building per-build telemetry, then designing bins from real percentile data. Same hardware, several multiples more headroom, no rewrite of any pipeline required.

An equities backtester built with the statistical honesty of a real-money system: selection-bias-aware deflated Sharpe with explicit trial count, stationary block-bootstrap confidence intervals, predicted-vs-realized calibration against a live trade journal, correlation-adjusted quarter-Kelly sizing with heat / sector / currency caps. The engine work (pure Polars strategies, vectorised regime detection with hysteresis, threaded scanner) is what makes the trust layer cheap enough to actually run every morning.

Replacing a legacy 121K-line per-SKU integer-programming procurement system, whose actual demand forecaster was this-year-over-last-year, with a two-stage decision engine: a LightGBM quantile demand forecaster feeding a HiGHS LP capital allocator. A four-way ablation cleanly attributes wins between the forecaster and the allocator, on a simulation engine that runs ~30× faster than the Python-idiomatic baseline and is locked by nine source-level invariants.

The same haversine kernel walked from a naïve pandas `.apply` through C++, Rust, Zig SIMD, and finally an analyzer-driven V7 in Zig that reads its own compiled assembly to land at 150 GB/s, plus a WebGPU compute lab in the browser. End-to-end demo of the optimisation work I do for clients.

Bare-metal Kubernetes on 4× Raspberry Pi 4 with Flux, Cilium, Tailscale, an in-cluster Zot registry, and MinIO. Hands-on platform engineering: the same GitOps patterns I apply to bigger clusters at work.