Here is a truth I have learned the hard way: most enterprise AI does not fail because the model is weak. It fails because the system around the model is naive.
A chat box on top of a PDF dump is not an agent. It is a party trick with compliance risk. Production AI in healthcare — or any regulated domain — needs the same rigor we demanded of payment APIs and claims pipelines a decade ago. Observability. Governance. Fallback paths. Cost controls. Human-in-the-loop where it matters. And above all, integration with the systems experts already trust.
What I build
My AI systems work sits at the intersection of distributed engineering and human expertise. I design platforms where:
- Agents are first-class services, not UI ornaments — concurrent execution, structured outputs, validation, tracing, and attribution for every invocation.
- Retrieval respects how experts actually read — hierarchy, structure, and relationships before flat chunks. (If you have read my Expert Vision research, you know why this matters.)
- Governance enables speed — tiered risk classification, reusable compliance patterns, and shared prompt libraries so teams move fast inside guardrails instead of around them.
- Failure is designed, not discovered — circuit breakers for model outages, cost-aware routing, PII detection, and incident playbooks that match how operators actually work.
The healthcare reality
At Centene, the stakes were never abstract. Agents needed to interact with claims, membership, and payment flows — real-time data, real latency SLAs, real audit requirements. That environment does not forgive architectural hand-waving.
So we built orchestration in Go. We wired agents into Kafka-fed operational context and governed lakehouse retrieval. We instrumented everything. We made deployment repeatable — from weeks of bespoke integration to days using shared primitives.
A question I keep asking
If an expert engineer sat down with your AI system tomorrow, would it amplify their judgment — or ask them to babysit a probabilistic intern?
That question keeps me honest. It is the standard I hold production agent platforms to. Not “does it demo well?” but does it earn trust under pressure?
The selected work below reflects that standard in practice.