Notes on building production AI systems — RAG, agents, and evals.
Jev: The Model That Doesn't Write Answers
TypeSafe's Jev returns typed decisions with calibrated probabilities instead of generating text. What a System One model is, the three primitives, what the request and response actually look like, and where the...
Do LLM Guardrails Actually Work? What My Own Numbers Say
An input gate, a model, an output gate — measured across 34 prompts and three runs: what a guardrail actually blocks, what it wrongly blocks, what it costs, and why running the...
Is RAG Dead? What My Own Numbers Say
A controlled test of RAG vs. long-context on the same corpus and the same model — accuracy, retrieval quality, latency, and real dollar cost, measured rather than assumed.
Production-Ready RAG Architecture: Core Patterns Explained
Why RAG demos break in production, and the patterns that fix it: chunking strategies, embeddings and vector stores, reranking and hybrid search, and how to actually measure retrieval quality.