I build things that make computers do the annoying parts of work for us — code generation, product research, video editing — and ship them faster than I probably should. Currently splitting time between AI infra at scale, two startups still finding their shape, and a slow-burn habit of fixing bugs in repos I don't own.

Comfortable in the space between "prototype" and "production." Not comfortable sitting still.

Fine-tuning a 14B code model to write infrastructure-as-code that actually passes lint, security scans, and terraform plan on the first try — turns out "sounds right" and "is right" are very different bars when the output has to compile. Most of my week is somewhere between reward-shaping a GRPO run and staring at an OOM traceback wondering what I did to deserve this.

When that gets too serious, I go argue with a RAG pipeline about which blender is objectively better, or find a bug in someone else's open-source repo and fix it uninvited.

LoRA · SFT · GRPO · hybrid retrieval · cross-encoder reranking · tree-sitter · RAG · XGBoost · LightGBM · AWS (S3, RDS, Lambda, ECR)

A running list of things I've poked at outside of work hours:

• fine-tuning models on hardware that was not built for fine-tuning models (RTX 3050, send help)
• a multi-agent product-research tool that argues with itself before giving you an answer
• someone else's subprocess matching logic that quietly matched the wrong PIDs — filed the PR, added the regression tests, moved on
• a personal EDR built on raw ETW telemetry, because watching syscalls fly by is a surprisingly good way to relax
• a short-form video tool that taught me more about DNSSEC than any of my CS courses did

Mechanical keyboards, forza, and open-world games — not always in that order of importance. I read changelogs the way other people read the news.