Highlights
- Built a TensorFlow scoring model and LangGraph agentic auditor (AWS Bedrock, Claude, knowledge-base grounding, human-in-the-loop escalation) that validated 50,000+ VA medical forms in 8 months, offsetting ~80 FTEs of manual review and sustaining the 92%+ quality score behind a 3%-of-invoice bonus
Project Narrative
On the VA MDE account, completed medical examination forms had to clear quality review before submission, and the stakes on both sides were high: forms the VA judged deficient were sent back as reworks the firm was not paid for, while the veteran’s compensation was delayed in the process. Review was manual and scaled directly with appointment volume. Quality also carried a direct financial incentive. Sustaining a 92%+ VA quality score triggers a bonus worth 3% of quarterly invoice value, so form quality maps directly to account revenue.
Phase 1 — Quality scoring model + automated feedback (PY25)
I built a deep-learning model in TensorFlow that scored submitted medical forms pass/fail with an associated quality score, trained on hundreds of thousands of historical exams labeled by their real pass/fail outcomes. The model worked from the forms’ free-text content. I paired it with a GenAI layer, Claude (3.5-era) via AWS Bedrock, that turned the model’s classification into specific, written pre-feedback delivered to the QA team, so reviewers had a lighter workload in writing and analyzing forms. High-confidence passes were auto-passed, removing manual validation for 1,100+ forms per month and feeding directly into sustaining the 92%+ quality score that earns a 3%-of-quarterly-invoice bonus. This number would scale up into ~5,000 a month on expansion.
Phase 2 — Agentic resubmission verification (PY26)
In addition to the POC classification model, the team expanded into an agentic auditor built on LangGraph to handle a second QA workflow. LangGraph was backed by knowledge bases of VA documentation and reference examples the agent could check against. The target use case was the resubmission loop: when the QA team flagged a deficiency and sent a form back to the provider, the agent verified whether the provider had corrected the specific issue the QA reviewer had raised. If the agent determined the issue was resolved, the form moved on; if not, it was escalated to a human reviewer. This allowed the QA team to concentrate scarce human attention only on the cases that genuinely needed it. This version validated 50,000+ medical forms in 8 months, automating an estimated ~80 FTEs of review effort, and formed part of the automation portfolio that drove ~$35M in margin on $170M+ in revenue.
The skill arc mirrors the system’s: Phase 1 was supervised learning plus a single GenAI call; Phase 2 was a stateful, multi-step agent reasoning over a knowledge base with a deliberate human-in-the-loop escalation path which was a meaningfully harder design problem.