Notes on building AI systems that survive production

Welcome. This is where I write about building AI systems that actually run in production, not just in a demo.

I spend most of my time on the parts of AI engineering that don't make it into launch tweets: keeping local and open-source models fast and reliable, wiring up retrieval that returns the right context, designing agents that fail safely, and building the data foundations that make any of it trustworthy.

What I plan to write about

• Local and open-source inference — running models with vLLM and llama.cpp, quantization tradeoffs, and when local beats a frontier API.
• Retrieval and knowledge systems — chunking, embeddings, reranking, and evaluating whether retrieval is actually helping.
• Agents — tool use, MCP, guardrails, and keeping agents observable when they act on real systems.
• Evals and reliability — how I measure quality before and after shipping, and how I catch regressions.
• The data underneath — pipelines, schemas, and the boring work that decides whether an AI feature is any good.

How this works

Each post is written in Markdown and lives in a single file in the repo, so publishing is just adding an entry and deploying. No CMS, no database, no lock-in.

More soon.