An NDA-safe, pattern-level account of production agentic systems built at VetClaims.ai. Private claim schemas, model choices, code, and pipeline details are intentionally omitted.

The Operational Problem

Veteran disability-benefits work brings together long records, clinical evidence, regulatory criteria, and multiple human handoffs. The cost of a weak automation is not just an inconvenient output: missing context or an unreliable handoff can create more work for operators and more delay for a veteran.

As Medical Innovation Officer at VetClaims.ai from August 2025 to May 2026, I owned AI initiatives from problem framing through delivery. The goal was to make complex workflows more consistent and useful without removing human judgment from high-stakes decisions.

The Agentic Approach

I built and iterated systems around reusable agent patterns rather than a single monolithic prompt:

Multi-agent coordination separated research, evidence review, drafting, and quality-control responsibilities.
Tool-calling loops let agents retrieve approved context and act through constrained workflow integrations.
Session memory preserved the right working context across longer tasks and handoffs.
MCP-style integrations gave agents a consistent interface to external tools without coupling every workflow to one implementation.
Human-in-the-loop checkpoints kept consequential decisions and final review with domain experts.

These patterns supported clinical-evidence pipelines, claims workflows, nexus-letter tooling, and adjacent operations systems. The exact implementation and proprietary data structures remain under NDA.

Reliability Before Demo Appeal

Production agents need to be judged on task completion, evidence use, controllability, and failure behavior—not just whether one response looks impressive. I treated prompt engineering and evaluation as engineering disciplines, with repeatable test cases, explicit review points, and feedback loops that informed each iteration.

The work shipped in two-to-four-week cycles while priorities changed quickly. That cadence required clean boundaries, reusable components, and observability that made failures understandable enough to fix.

Outcome

The systems replaced manual handoffs with AI-enabled workflows across benefits and operational domains while preserving human review where the stakes demanded it. I remained hands-on from architecture through delivery, working across engineering and product responsibilities.

What I Can Discuss

I can discuss the system-design patterns, evaluation approach, human-review strategy, delivery tradeoffs, and lessons from taking agents beyond prototypes. I cannot share private code, claim schemas, model names, customer data, or confidential architecture diagrams.