The interesting part of AI engineering is everything underneath the model call.
I build with a fleet of coding agents, and when a layer under them breaks or falls short, I fix it upstream.
Now: vidtheque, a self-hosted video knowledge base my agents query mid-task, and voice for those agents in localvoxtral.
Selected work
- vidthequePython · JavaScript
The talks you don’t have time to watch, turned into knowledge an agent can query mid-task. Every answer carries its receipt: the sentence, the slide, and the second it happened. Self-hosted, on hardware you already own.
- localvoxtralSwift
Realtime, fully local dictation in the macOS menu bar: words appear while you’re still speaking. Built for prompting coding agents by voice, so it joins the exact Claude Code session under your cursor and polishes against what is on its screen.
- working-setPython
How many agents a given GPU configuration keeps warm, and which constraint binds first: KV cache, decode bandwidth, or prefill compute. A scenario model, and an explorer that answers with a verdict and the vLLM flags to match.
- fastcontextPython
Read-only repository-exploration agent that coding agents delegate to over bash: cited file:line answers, not context noise.
Upstream, where the layer underneath was the problem: mlx-audio-swift (the Voxtral realtime streaming path), llama.cpp (thinking blocks surviving the server’s message conversion), mlx-lm (two prompt-cache correctness bugs), Node (a V8 segfault on small-ICU builds, with the fallback shipped the same day).