Agentic AI engagements
Agents that survive production, not just the demo
I have built two agentic systems inside a regulated healthcare platform — a clinical summarisation engine handling 1.4M notes a month, and a spec-driven development workflow adopted by 35 engineers. Both had to pass compliance, cost review and on-call. That is the standard I bring to your build.
A four-week path to something running
Which steps a model should own, which stay human, where the data actually lives, and what "correct" means in numbers.
Tool contracts, state and retry semantics, an eval set from your own records, and the cost ceiling written down.
One workflow, end to end, on your infrastructure — queues, observability, audit trail, feature-flagged rollout.
Regression gates in CI, runbooks, dashboards, and a written plan your team can execute without me.
What I bring that a prompt engineer does not
- Sixteen years of distributed systems — the failure modes of agent orchestration are the failure modes of event-driven systems with worse error messages.
- Kubernetes and AWS operations, hands-on: EKS, Lambda, ECS, Terraform, and the ECS-to-EKS migration to prove it.
- Regulated-data experience: HIPAA, PII boundaries, audit trails, HL7 FHIR and SMART on FHIR integrations.
- Delivery ownership — scoping, staffing, timelines and stakeholder communication, from five years managing teams.
Typical engagements
Retrieval, permissions, citations, and an eval suite so the answers stay right as the corpus grows.
The pattern behind Chartwell: extraction, structured output, reviewer workflow, measurable time saved.
Long-running, multi-tool processes with durable state, compensation and human approval steps.
Spec-driven development installed team by team, with the cycle-time measurement to show it worked.
Start with the architecture sprint
Two weeks, fixed price, ends in a build plan you own.