Independent Research
01 · Model benchmarking on real bugs
Which LLM can repair a real historical bug? Ten models × 13 bugs from this project's history, the test suite as judge, all 130 transcripts kept. A free local 120B repaired 10 of 13; two silent-logic bugs (a dropped await, a dead guard) defeated every model until the newest frontier tier — one model finally repaired all 13.
read the A3 →02 · SRTM → GRIT world generation
Public elevation data to governed in-game terrain, with the pipeline that refused to fake a lake.
read the A3 →03 · CI and A3 error reporting
A 0.4-second inner loop, a mutation-tested suite, and Toyota-style A3 reports an agent can act on.
read the A3 →The world map: Mediterranean terrain generated from satellite elevation data, rivers earned from a rainfall simulation (article 02).
Alongside my engineering career, I build Race to the Iron Age — a solo-developed strategy game spanning a world map generated from real satellite terrain, a turn-based tactical hex layer, and real-time battles. It is deliberately too big for one person: the project is my working laboratory for how far AI language models can multiply a single engineer's output — and how to keep the result trustworthy. Everything below is measured on this codebase, not toy examples.