Answers from your protocols, with the page attached.

• 01

  Protocol and formulary lookup

  Staff ask in plain language and get the current protocol with its revision date, rather than opening the shared drive and hoping the file name is honest.

• 02

  Payer policy retrieval

  Coverage criteria differ per payer, plan and year. Retrieval scoped to those dimensions stops one payer's rule being quoted at another.

• 03

  Patient-education drafting

  Discharge and pre-procedure instructions drafted from approved material at the reading level you specify, in the patient's language, for a clinician to approve.

• 01

  PHI does not go in a prompt by default

  We de-identify at the boundary and re-associate after, so the model sees the clinical question without the identity attached. Where full context is unavoidable, it runs under a BAA or inside your own infrastructure.

• 02

  The system drafts; a clinician decides

  Nothing reaches a patient without review on any pathway that touches clinical content. That gate is a hard architectural boundary, not a configuration toggle someone can turn off later.

• 03

  Refusal beats a plausible answer

  Out-of-scope questions get an explicit hand-off to a human with a callback, not a hedged paragraph. We test the refusal path as carefully as the answer path.

1. 1

   Data inventory

   What docs / databases / APIs / feeds matter? What's PII vs public? What changes daily vs quarterly? We map it before writing code.

2. 2

   Retrieval architecture

   Hybrid (BM25 + vector) retrieval, chunking strategy, reranking, citations. Tuned for your domain not a generic benchmark.

3. 3

   Eval suite

   Real-world question set with golden answers. Faithfulness, context precision, latency. So you know when changes ship a regression.

4. 4

   Guardrails

   Prompt-injection defense, PII scrubbing, refusal patterns, output schemas. Especially critical for customer-facing deployments.

5. 5

   Production + iterate

   Cost-effective inference (caching, fallback models), monitoring, weekly eval reports. Fine-tune when the data justifies it.

• Epic & FHIR APIs
• Athenahealth
• HL7 v2 feeds
• Twilio & WhatsApp Business
• Zendesk
• Snowflake
• S3 with encryption at rest
• Practice-management and RCM systems

FAQ

RAG for healthcare and clinical documents: your questions.

• How do you stop it answering from general medical knowledge?

  The prompt and the evaluation both treat an uncited claim as a failure. If retrieval returns nothing relevant the system refuses and offers a human — and that behaviour is a scored line in the eval suite, so a model swap can't quietly erode it.

• Can it read scanned faxes and PDFs?

  Yes — OCR plus layout-aware extraction, which is most of the work in practice. We report extraction accuracy per document type up front, because a fax from 1998 and a structured PDF are not the same problem.

• Should I use RAG or fine-tuning?

  RAG for facts that change. Fine-tuning for style, format, or domain reasoning. Most production systems use both — and we'll tell you which mix is right after a 1-week discovery.

• Can I use my own LLM (open-weights)?

  Yes. We work with Llama, Mistral, Qwen, and similar — useful for data residency, cost predictability, and air-gapped environments.

• AI agents for healthcare operationsAgents that handle scheduling, reminders, insurance verification and document chase across voice and messaging — escalating anything clinical to a person immediately.Read
• Healthcare AI governance and readinessUse-case triage against clinical risk, a data-handling design that survives a HIPAA or DPDP review, and an oversight model your clinical governance body will actually accept.Read
• 30-day AI sprint for healthcarePick the administrative workflow that hurts most. In 30 days you get working software, an eval score, and an honest recommendation — including the recommendation not to proceed.Read

• RAG for banking and financial servicesRetrieval over circulars, product terms and internal procedure — returning the clause, the revision date, and a link to the page it came from.Read
• AI product search and catalogue RAGSemantic search and grounded product Q&A over your catalogue, reviews and size guides — with stock and price read live rather than remembered.Read
• RAG for SaaS productsRetrieval over your docs and your customers' workspace content, with isolation enforced in the index and tested adversarially.Read