Research, competitions, and infrastructure that exists because I wanted the experiments to be repeatable.
What changed because the project existed — not just what technology it used. Links go to the result, the live thing, or the lab page that describes it.
Part of Team Reki, which placed second in ARC-AGI-3 Milestone Prize #1: an interactive reasoning benchmark built to defeat memorisation. My lane ran on the home lab’s GPUs — fine-tuning and evaluating open models against novel rule families, with the experiment scaffolding, checkpoints and logs kept reproducible across machines.
Competitive machine learning as a forcing function for rigour. 18th of 2,963 in the 2026 NeuroGolf Championship; 78th of 4,729 in Orbit Wars; 123rd of 4,182 in NVIDIA’s Nemotron Model Reasoning Challenge; 309th of 4,186 in AI Agent Security: Multi-Step Tool Attacks.
A single web app that turns a plain-text infrastructure inventory into a live map: ping sweeps of every subnet, DHCP-lease discovery of unknown devices, service checks, SLURM queue and GPU telemetry, hypervisor state, a maintenance calendar, rendered docs, and derived alerts. Python, asyncio, SQLite; agents on each host push metrics every 30 seconds.
SLURM across two RTX 4090 workstations and an arm64 NVIDIA DGX Spark: configless nodes, cgroup-isolated GPUs, separate architecture partitions, accounting, and a desktop that only takes work when nothing else is free. JupyterHub sits beside it with a per-GPU profile picker over a shared NFS environment, so the same hardware serves batch sweeps and notebooks.
A web application for a small accessories company: product catalogue, vendor bills, expenses and invoices on PostgreSQL, with a reviewed release process (candidate build, restore-from-backup reconciliation, then cut-over). Public catalogue at pkaccess.com; the business portal stays private.
Web, API and monitoring stack for keeping track of tools and who has them. Self-hosted behind the lab’s reverse proxy with its own certificates and metrics endpoint.
One slot stays open on purpose. It signals that the collection is alive and gives unfinished work somewhere to land.
ARC-AGI-2 generators: a corpus of 3,889 ARC-AGI-2 puzzles, each solution written as a small Racket program and paired with a Python generator that produces fresh inputs.