Onboarding in 2.5 weeks. Compliance answers in 30 seconds.

The challenge

Placing clinical professionals — nurses, allied health workers, locum physicians — involves navigating a dense web of regulations: CMS guidelines, state-by-state licensing reciprocity rules, facility-specific credentialing requirements, and the agency's own compliance policies. Getting this right matters: a placement error is not just a business problem, it's a patient safety issue.

New placement coordinators typically took six weeks to reach the point where they could handle most matters independently. During that period, they were generating 40+ escalations per day to senior coordinators — questions about licensing rules in specific states, edge cases in the credentialing process, interpretation of CMS requirement changes. The senior team was answering the same questions repeatedly, spending hours each day on ad-hoc coaching rather than strategic placements.

The agency's director of operations wanted to improve onboarding speed and reduce the senior team's interruption load — but without creating a rigid FAQ system that couldn't handle the nuance of real compliance questions. The rules change frequently, and coordinators needed answers they could trust, not just a static document they'd have to interpret themselves.

The solution

Gilligan Tech built an internal knowledge base using a RAG pipeline over the agency's full compliance document set: the employee handbook, CMS guidelines (current and archived), state licensing rules for all operating states, facility credentialing checklists, and a structured database of historical placements with outcome notes.

The knowledge base is queried through a conversational interface available to all coordinators. A coordinator can ask: "Can a licensed practical nurse in Texas work a locum shift at a Kansas SNF under CMS conditions of participation?" — and receive a grounded answer within 30 seconds, citing the specific guideline sections that apply.

Gemini 1.5 Pro's long-context capability was essential: many compliance questions require reading multiple regulation sections simultaneously. The model can hold the full CMS CoPs plus relevant state rules in context when answering a complex cross-jurisdiction question — something a keyword search or simple FAQ cannot do.

Implementation

1. Knowledge audit: All compliance documents were catalogued — 847 source documents including handbook sections, CMS guideline chapters, state licensing rules for 18 operating states, and 3 years of placement records. Documents were tagged by type, jurisdiction, and last-updated date.
2. Ingestion and chunking: Documents were chunked at section boundaries with metadata preservation (guideline number, section title, jurisdiction, effective date). Placement records were structured into a separate retrieval corpus.
3. Embedding: AWS Bedrock Titan Embeddings V2 generated embeddings for all chunks. A separate embedding index was created for the placement history corpus to allow precedent-style queries.
4. Query pipeline: User queries retrieve the top-ranked chunks from both the regulatory and placement corpora. Gemini 1.5 Pro synthesises a grounded answer with citations. Confidence is estimated and flagged when the query falls into a novel edge case.
5. Update workflow: When regulations change, affected document sections are re-chunked and re-embedded automatically. The operations team receives a weekly digest of queries that triggered low-confidence responses — flagging knowledge gaps for human review.

Results

Technology