The AI trio — govern, build, deliver

This is the proof behind the AI & automation positioning: not slideware, but working code and reference architectures you can read. Three sibling repositories form one coherent stack.

The story

• Govern & route — sovereign-llm-gateway. An LLM gateway with per-agent cost and budget enforcement, vendor abstraction, a local-model fallback for sovereignty, and Prometheus observability. This one runs end-to-end (docker compose up).
• Build — sovereign-copilot. A reference architecture for a trustworthy agentic product: deterministic tool contracts (MCP), retrieval grounded in your data, L1–L4 evaluation gates with goldens, and answers that trace to a recorded call chain.
• Deliver — maestro (MIT). A reference architecture for spec-driven delivery: agents propose, humans dispose, with functional and technical gates enforced through GitHub branch protection.

The honest part

One of these runs end-to-end today; the other two are reference architectures. They are not three production systems, and I won't present them as such. That honesty is the point — it's the same discipline I bring to a client's AI programme.

Why it matters

Together they map onto how I run AI delivery for real: govern the models, build a product you can trust on top of them, and deliver software with agents under human control.

→ See the repositories on the portfolio (links open once the public-surface prerequisites close).