Research engineer in generative vision, five years in. Most of that time has gone into instruction-based image editing with diffusion models — getting them to understand what to change, touch nothing else, and run fast enough to ship. Off the clock, I apply the same principle to myself: anything an agent can do, I don't do twice.
Controllable image editing — instruction-driven local edits that preserve identity and layout; portraits, hair, and color are where "almost right" still looks wrong.
Distillation & acceleration — multi-teacher → single-student on-policy distillation; compressing dozens of sampling steps into a few without losing quality.
Evaluation — "did the edit work" can't be eyeballed; identity baselines, paired comparisons, and reproducible eval pipelines.
Agentic tooling — wiring training, image review, bookkeeping, and scraping into LLM agents. Built for myself, open-sourced along the way.
- Baseline before claims — without a "leave it unchanged" control, every "improvement" is an illusion.
- Data sets the ceiling: a week cleaning data beats a day tuning hyperparameters.
- It's done when it reproduces locally; every automation runs on a smoke set first.
- Writing down what broke is worth more than writing down what worked.
PyTorch · Diffusers · LoRA · FSDP / DeepSpeed · CUDA · Docker · Linux · Claude Code / MCP · n8n · Python
Less, but better.


