Writing
Technical posts on the hard parts of what I've built.
Why I modeled LLM rate limiting on TCP congestion control
June 2025Most LLM wrappers hard-code rate limits or use exponential backoff. I built a controller that discovers and saturates the actual provider ceiling automatically.
Map-Elites for adversarial prompt generation
July 2025How I used a quality-diversity algorithm to generate diverse adversarial test suites for LLM behavioral norms — and why random sampling wasn't enough.
Multimodal RAG over clinical guidelines: what the benchmarks don't tell you
September 2024Building a RAG system over 10K clinical guideline chunks taught me that retrieval quality metrics and real physician utility are measuring different things.