Things built end to end. Some shipped, some were research questions that needed answering, one is a football game. Most are open source.
An LLM inference server built from scratch to study scheduling: naive generation, then static batching, then continuous batching, then a paged KV cache — each step measured rather than assumed. Continuous batching reached 278 tok/s on a T4, 9.6× the naive baseline, with no custom CUDA kernels. The interesting part is the measurement discipline: published optimisations are audited against a real noise floor, and results report goodput under a latency budget instead of headline throughput. That surfaces the unflattering finding that the fastest engine misses its SLO once the queue backs up.
A harmonised analysis across eight public resting-state EEG datasets — 504 participants, eight clinical conditions plus controls. One preprocessing recipe and one global microstate model fit on pooled peaks, so the maps are identical for every subject and features are actually comparable across sites. Then classification against a band-power baseline, with cross-site external validation, batch harmonisation, age-confound checks and permutation tests.
A Shopify app for a deadline. On 26 August 2026 Shopify stops running merchants' legacy checkout scripts, and purchase tracking sitting in that box goes quiet with no error — the numbers just flatten. Holdfast reads what's there, rebuilds each tag as an official Web Pixel, and adds a server-side leg that reports every paid order from the store rather than the browser. Both legs share one event ID, so an order is never counted twice.
A research harness built to kill its own ideas. Around 70 registered experiments, 15 strategies falsified — including the one that looked best — and a single survivor at 0.85 walk-forward Sharpe across a 77-instrument cross-sectional ranking, positive in 12 of 15 years, net of costs and out of sample with trial counts attached. It publishes to a daily research terminal that scores the model against its own past predictions. The falsifications are the point.
Three prediction-market mechanisms that cannot be backtested, because each needs live order books, forward outcomes, or reward accrual. So they run forward on paper instead: a neg-risk arbitrage executor that prices every opportunity against real depth, a shadow book that copies top wallets and grades at resolution, and a maker simulator tracking adverse selection. Ledgers are committed on every run, wrong answers included.
A family chore app. Kids complete tasks and multi-step routines to earn points and spend them in a reward store their parents curate; parents get charts, per-routine stats, and a push notification when something's done. Built to ship — auth, realtime sync, an access-code gate, and store build configuration for both platforms.
A World Cup fantasy-draft manager in a single file with no dependencies and no build step. Draft against a £100m budget, then take your squad through a 32-team tournament against AI managers: minute-by-minute simulation, xG, shootouts, a tactics board that punishes players out of position, stamina, suspensions, injuries. Open the file and it runs.
A job search run as a pipeline. The extension autofills applications from a saved profile and drafts free-text answers, then a dashboard does the rest: a morning briefing of postings from the last 48 hours, 0–100 match scoring, target companies watched directly through their Greenhouse and Lever APIs, and a tracker that logs every form you fill. The writing tools include a resume tailor that outputs an explicit list of things it could not claim truthfully — the honest half of the feature.
Control software for an underwater competition robot: a desktop interface for thrust and multi-camera views, a seagrass-detection vision module, and the Arduino firmware underneath including a neutral-buoyancy mode. Hardware you can drown, which changes how you write the software.