AI architect, builder, and educator. I connect systems thinking with working software — from context compilation and local-first AI to privacy tools and a storybook game portal built with my daughter.
The layer between what an AI system can retrieve and what it actually reasons with is a compiler problem, and most architectures never made it explicit. I publish that layer as context compilation.
- contextos — Context IR object model, JSON schema, lowering stage, and the CompileBench specification. Apache-2.0.
- Papers with DOIs: context compilation · eight metrics — archived snapshot · compiler passes · paged context memory · quantized context
- Frontier-Bridge — memory-hierarchy planning for giant MoE models on hardware you already own. Every published number is measured or absent; where the numbers don't work, the runbook says so.
- ai-infrastructure-reality-ledger — keeps infrastructure proposals, construction, and operating capacity as separate facts, each attached to evidence. Live demo — public beta running a synthetic demonstration corpus.
- data-broker-eraser — opt-out automation across 20 people-search brokers, re-run every 30 days because a one-time opt-out is theater. Runs on your machine; your data never enters an AI assistant's context. The README publishes the honest results table, including the brokers that blocked us.
18 courses on Pluralsight, 2023–2026, from RAG and LLM agents to AI-assisted development at team scale. Companion repositories below. → Full course index
- kids-arcade — a no-accounts, no-tracking game portal for my daughter, self-hosted on our home lab. Play it.
How I work: claims carry evidence or say they don't. Local first. Apache-2.0 / MIT. Writing → brianletort.ai