A 24/7 pipeline that polls Dublin's GTFS-Realtime feed and classifies every scheduled Dublin Bus / Go-Ahead trip into one of six honest outcomes, grading its own uptime in public alongside the routes it measures. A feed-health gate caught a real NTA data outage on its first live day and withdrew that day's verdicts rather than publish false accusations. Paused since July 2026 on the same suspended VM as the bikes forecast, with the 14-day baseline frozen at day 4 and the gap shown on the site.
A month of one graduate’s real job search and AI-assistant operations, parsed from messy markdown trackers into a queryable DuckDB database and published as an anonymised static dashboard. A privacy gate enforced in CI was adversarially attacked and closed on 7 real leak paths before anything went public, behind a 164-test suite. Nine SQL analyses over 50 tracked applications and 65 logged assistant operations, deliberately framed as descriptive rather than inferential at that sample size, every rate published beside its raw numerator and denominator.
Pre-registered Monte Carlo projection of young footballers’ careers on 24,057 FBref player-seasons, cross-checked on unseen 2025–26 data at six horizons. Verdict: NOT PROVEN, skilled but under-confident.
A personal AI butler that runs my mornings – now open source. A voice-controlled Iron-Man-style HUD (Electron, 100% local speech-to-text via whisper.cpp) sits on top of a scheduled, fully unattended agent pipeline that delivers a grounded morning brief to Telegram at 08:30, with two-way remote commands, DPAPI-encrypted credentials, and fail-closed safety rules. Elite interview stories – including a shipped-and-fixed command-injection bug – but no 2-minute demo: it is local, personal, and scheduled.
A weight-cut platform for judo and BJJ athletes, built by a 16-year black belt for his own sport. Live with alpha users: streaming AI estimates, Upstash Redis rate limiting, Resend email automation, an installable offline PWA, full technical SEO, and GDPR-compliant export and deletion.
Pre-registered backtest over 2,418 resolved Polymarket markets: favourites win 90.6% of the time and still lose ~1% after fees. Verdict: NOT PROVEN.
Final-year Pine Script v6 strategy for NQ1! E-mini futures using Inverse Fair Value Gaps, Change in State of Delivery, and a liquidity-sweep filter, with fixed risk rules and honestly documented out-of-sample caveats. Its own sequel, the Quant Strategy Research Program, put this edge on trial – and disproved it.
Monte Carlo stress-test of a real 72-trade NQ futures record – bootstrap resampling, order reshuffling, and drawdown-risk distributions, with an honest writeup of what survives.
230+ commit full-stack trade journal and market-research app – Next.js dashboard, authenticated trade logging, analytics pages, a Python/FastAPI back end, tests, schemas, and Vercel deployment.
Python engine with a strategy registry, slippage/commission model, and walk-forward split. Benchmarks SMA Crossover and RSI Mean Reversion against the FYP strategy; GitHub Actions auto-updates results.
