Work

I've been working on AI infrastructure.

Local model runtimes, hardware compatibility, and the tooling that keeps the work honest. Four areas, roughly in the order I touch them.

Local AI runtime

C++

Contributing to an open-source local LLM server: server internals, the CLI, model management, and hardware backend support, including getting GPU acceleration working correctly on consumer Radeon cards.

A runtime sized to one machine

C++ / CMake / Python

A reduced build of that same kind of server, sized for constrained GPUs, with build-time selection that compiles out backends the hardware can't run.

Benchmark and evidence tooling

Python

Unattended, resumable model benchmarking that survives a crash and keeps a record of what it measured and why, so a result can be checked later rather than just trusted.

Work triage system

Django / SQLite

A local tool that scores incoming work by feasibility and how well I actually know the code it touches, so unfamiliar areas get more proof and less guessing before I start.

I write about a lot of this as I go.

Read the writing →