I build machine learning systems that ship: match-forecasting models, councils of AI agents, and spatial models that fix real map data. I care about honest evaluation as much as the model itself.
A Premier League match-forecasting product, from raw data to a live site.
Ingests 33 seasons of match and player data, builds leakage-safe rolling pre-match features, and serves calibrated home, draw, and away probabilities. The production model is Dixon-Coles with ridge-shrunk team ratings and time decay, and an XGBoost blend is only promoted when a gameweek-block bootstrap confidence interval says it actually wins. Users can sign in and save lineup scenarios to re-run against the current model.
matches in final fit
12,809
walk-forward accuracy
46.6%
log loss
1.028
RPS
0.209
log loss falls across model iterations (illustrative)
Three AI agents with competing mandates negotiate a live production incident.
One agent restores service, one contains harm, one protects customers, and each sees only the evidence it's entitled to. They propose blind, rank options with a Schulze vote, critique and amend, and act alone only on reversible steps they agree on. When they genuinely disagree about something irreversible, MAGI phones a human with the crux. I built an evaluation harness on 13 decision points adapted from real postmortems (Cloudflare, GitLab, AWS S3, Knight Capital) and reported the results honestly, including where the council lost to simpler baselines.
Co-built ClawForge, an AI agent factory for small businesses: describe a need in plain English and get an agent with tools, memory, guardrails, and a deployment plan. Built on NVIDIA Nemotron and Brev.
epoch 04 · Jan – Mar 2026
Research Overture Maps Foundation
TruePin: an XGBoost classifier on Uber H3 spatial features to correct map POI entrance coordinates, beating the Overture baseline by 7%. Presented to engineers from Amazon, Meta, Microsoft, and TomTom.
epoch 03 · May – Jun 2025
Software engineer LAIQA
Shipped a real-time biometric tracking module on Apple Watch and Fitbit APIs, containerized MongoDB services, and contributed to an HRV-based stress detection algorithm.
epoch 02 · Apr – Jun 2024
Developer intern Here For You
Worked on a platform built to support people going through a tough time.
epoch 01 · Sep 2023 – Apr 2024
Software engineer Soshal
Full-stack developer on a Flutter app for UCSC students to discover club events, from MVP to university pitch.