← All work  ·  Case 07 of 09 · POC

Enterprise Healthcare GRC SaaSHealthcare ComplianceSolo · Product + AI2026 · POC

Five small agents. One rule: never fabricate a citation.

A multi-agent pipeline that watches regulatory sources, scores relevance per customer, and drafts a grounded impact assessment, with mandatory human approval before any of it counts.

<15%
Target edit-rate
AI drafts a human must rewrite
5
Single-purpose agents
ingest → relevance → detect → synthesize → Q&A
0
Fabricated citations tolerated
grounded-or-silent by design
$0
Infra cost
runs offline, mock-AI demo mode
01The problem

Missing a regulatory change isn't a small mistake.

Healthcare compliance teams manually watch a shifting body of federal and state regulation. Today that means reading the source, noticing a change (or not), judging whether it applies to their facility, and writing an impact memo from scratch, every time. Miss one and the consequence is a failed survey or a repayment demand, not a missed Slack message.

02Discovery

Detecting change is easy. Judging relevance is the real cost.

Diffing a document is trivial. The expensive human work is deciding whether a change actually applies to a specific facility, and then translating dense regulatory text into a plain-language, correctly cited recommendation. Relevance has to be scored against an explicit customer profile (programs, facility types, jurisdictions), not judged globally, or the signal drowns in noise.

"A compliance tool that fabricates a citation once is dead. Over-conservative beats confidently wrong."
03Approach

One job per agent. A human decides everything.

Instead of one large prompt trying to do everything, the pipeline splits into five single-purpose agents: ingestion detects and diffs changes, relevance scores each one against a customer profile (acting as a cost gate before the expensive step), detection classifies severity, synthesis drafts a cited impact assessment grounded only in retrieved source text, and a conversational agent answers ad-hoc questions the same way.

The non-negotiable rule: the synthesis agent cites retrieved source text or says "source not identified." It is never allowed to guess a citation. Every assessment then lands in a human approval queue; nothing publishes unreviewed.
04Outcome

A working pipeline, honestly labeled as a POC.

This is a solo proof of concept, not a production deployment, and I'm explicit about that rather than overstating it. It runs fully end to end, including an offline mock-AI mode so the whole flow is demoable with zero cloud credentials. The design target is a sub-15% edit rate: the share of AI-drafted assessments a human has to meaningfully rewrite before approving, the single number that proves whether the AI is actually trustworthy or just theater.

05My role

What I owned, solo.

Architecture
Designed the five-agent pipeline and the model-tier abstraction that lets any agent be upgraded later without touching agent code.
Guardrail design
Built the grounding-only citation rule, the advisory-not-gating validation, and the mandatory human approval queue.
RAG
Chunking, embeddings, and retrieval feeding the synthesis agent's grounded context.
Full build
Backend, frontend, and a dual local/production runtime so the demo needs zero setup.
06Lessons

Design for the acceptable failure mode, explicitly.

In high-stakes AI, guardrails are the product. Human-in-the-loop isn't a fallback bolted on afterward; it's the reason the tool is adoptable at all in a domain where being confidently wrong is worse than being slow.

A single trust metric beats a dozen vanity ones. Edit rate is a direct proxy for "is this actually good enough to use," and designing the whole pipeline around moving that one number kept the scope honest.

Let's build what actually ships.