Blog
Research notes and build logs. The newest ones are interactive, with the real numbers from each run.
Resource Optimization
Profiling a 16-stream YOLO pipeline on one laptop GPU with Nsight Systems: from 27 seconds to 7, mostly by asking what each thread was waiting for.
Optimized Multimodal Support
Giving a text-only Llama-3.2-1B grounded vision and timestamped hearing by training nothing but a connector, on one 8 GB laptop GPU.
Building Keystrike: a real-time typing race with Socket.io
How I built a multiplayer typing speed game in TypeScript, and the Socket.io quirks that came with it.
Building LLamaPG: RAG as a service with pgvector
How I built a containerized RAG platform using FastAPI, PostgreSQL with pgvector, and LangChain, and what I learned about vector search performance at scale.