Deep-dives on AI engineering, production ML, and the systems I build. Written for engineers, by an engineer.
The hardest part of building an NL-to-SQL system isn't the language model — it's reliably conveying the schema to the model without hitting context limits or leaking sensitive data.
Running an LLM in production without observability is like flying a plane with no instruments. Here's the architecture I built to give AI teams full visibility into every inference.
Most RAG implementations treat retrieval as an afterthought. I spent three months obsessing over it. Here's what I learned about building retrieval that is fast enough to feel native.