written up

Case studies

Back to portfolio

Three systems from the lab, written up: what the problem was, what I built, and where the limits are. The AI on the homepage can answer questions about any of them.

An AI-native portfolio on self-hosted infrastructure

Why this site is a working system instead of a static resume. RAG chat over a curated corpus, a local model, deterministic guardrails, and a redaction layer, all running on my own stack.

Next.jsQdrantOllamaRAGMCPAuthentikCaddy

A multi-agent operating model that ships real work

Routing, shared memory, receipts, and human approval gates for a lab where Claude Code, Codex, Hermes, and local models deliver work as a coordinated team.

Claude CodeCodexMCPorchestrationPythonSQLite

Outcome-based evaluation for AI agents

Grading agent work by user-visible outcomes instead of activity. Separate goals, counter-metrics, no composite score, and an evaluator that fails closed.

evaluationgovernancecounter-metricsPython