Defra serving now
The current engagement (via Equal Experts): building the digital service for extended producer responsibility for packaging, how the UK tracks reprocessed and exported packaging waste. Government-scale delivery, live right now.
Today's Special
Hard problems, eaten for breakfast.
Thirty years of shipping well-tested software, fast. And, unusually: building and operating real multi-agent AI systems, not just using the tools. Senior contract engineer. Reading & remote.
The Full English
The whole track record, all the trimmings. Recent work via Equal Experts.
Parloa agentic AI, in production
Parloa's AI Agent Management Platform puts agentic AI on real contact-centre calls, at scale, in any language. The engagement took it from proof-of-concept to production: a custom AI orchestration framework, and thousands of automated tests to keep LLM non-determinism honest.
My plate was developer experience: AI-tooling onboarding for the engineering teams, template and GitOps automation, GitHub org administration, and a Datadog → OpenTelemetry migration.
Also recently: Kaluza · News UK
And a deeper back-catalogue: Contingent, Hargreaves Lansdown, Tesco, Just Eat.
The unabridged menu (a proper CV, with the lot) is available on request.
Off the menu — free refills
Open source: free, community-fed, help yourself.
The Kitchen
Round the back, the robots run the night shift.
My home platform is built like production, because it is one: upstream Kubernetes on Talos Linux, Cilium (eBPF, Gateway API), ArgoCD GitOps, Sealed Secrets, cert-manager, CloudNativePG, Authelia SSO, and full observability with Prometheus, Grafana, Loki, Tempo and OpenTelemetry.
And it is operated by a fleet of AI agents I built and direct: monitoring, alert triage, routine maintenance. Not a slide, not a demo: a live, in-production example of agent-driven operations, on shift every night.
If you're wondering whether agent-native delivery is real, this is the kitchen you ask to see.
The Recipe
How everything gets cooked.
Ingredients
- Thirty years of delivery
- A disciplined functional style
- Domain-driven design
- The right tests at the right levels (seasoned to taste, not smothered)
Method
- Parse, don't validate: make illegal states unrepresentable before they reach the pan.
- Model the domain, then let the types hold the shape.
- Ship small and often; keep the plates moving.
- Review everything the agents cook. Every plate leaves the pass checked.
AI-native means more rigour, not less.
Table for one?
Serving since 1995.
Graeme Foster, trading as Code For Breakfast Ltd. Based in Reading, working remote-first across the UK. Outside IR35. Recent engagements via Equal Experts.