home
06writing
Writing
Notes on agentic AI, evals, infrastructure, and shipping LLM systems in production.
Should we switch models? A tool that answers Ship or Don't ship
How upgrade-report uses paired statistics, a noise floor and non-inferiority testing to decide whether an LLM app should move to a new model or prompt.
llmevalsstatisticspythonciThe LLM Proposes, the Code Disposes: Building a Private Search Engine for My Photos
How I built natural-language search for my own photo library that runs entirely on one home machine, and why a pile of regexes beat a 14B LLM at understanding queries.
machine-learningsearchclippgvectorllmself-hostedWriting an LLM inference engine from scratch in Rust
And making it faster than Ollama on my CPU
LLMLLM InferenceRustAIGenerative AICPU OptimizationSIMDQuantizationTransformerAttentionKV CacheSpeculative DecodingOllamallama.cppLocal AIRAG Evaluation Strategies That Actually Matter
Retrieval and generation get all the attention. Evaluation is where you find out if any of it actually works — here's how to build that layer properly.
RAGEvalsLangSmithRAGASRetrievalLLMAI EngineeringProduction AI