AI & automation

This is a peer of the other four areas, not a bolt-on. The same discipline that makes an integration platform trustworthy — contracts, observability, governance — is exactly what most AI projects are missing. I build agentic systems the way I build any production system: with cost controls, evaluation gates, and an audit trail.

What I do

• Agentic systems with guardrails. Model routing, per-agent cost and budget enforcement, vendor abstraction, and a local-model fallback for sovereignty — with Prometheus observability so spend and behaviour are visible.
• Retrieval that's grounded. RAG with proper embeddings and reranking so answers come from your data, not the model's imagination.
• Evaluation as a gate. L1–L4 test layers, LLM-as-judge, and golden datasets so a regression is caught before it ships — the AI equivalent of acceptance criteria.
• Governance you can audit. Deterministic tool contracts (MCP), answers that trace to a recorded call chain, and an "agents propose, humans dispose" model enforced in the delivery pipeline.

The proof

I don't just advise on this — I build it. The portfolio is three sibling repositories told as one govern → build → deliver story. In the interest of honesty: one of them runs end-to-end today; the other two are reference architectures. They are not three production systems, and the cards say so plainly.

Evidenced by

• The AI trio — the govern→build→deliver case study behind this area.
• Portfolio — the repositories themselves, with honest maturity labels.

Background: DeepLearning.AI Agentic AI (2025); 20+ years building data and ML systems.