Amir Omidvar
← All work

Applied AI · Healthcare · 2025 to 2026

Clinical knowledge platform with a grounded AI assistant

An assistant clinical staff can trust, because it answers only from approved manual text and a second model checks it.

Ask the Manual
What are the onboarding steps for a new team member?

They complete identity setup, then role orientation and system provisioning before their first day [§4]. Badge access follows on day one [§9].

✓ verified by 2nd model§4§9
Ask a question…↑

Representative UI, abstract, with no client data.

Second-model
independent grounding check gates every answer
Verbatim
document imports splice source text; the LLM structures, never rewrites
De-identified
no PHI; strictly grounded in approved content

How it works

1Approved manual

full document cached as context; bounded corpus, no RAG miss

↓
2Answer model

Sonnet drafts an answer with inline section citations

↓
3Verifier model

Haiku independently checks every claim is supported verbatim

↓
4Gate

unsupported claims flagged; refusal detected when the manual does not cover it

↓
5Human triage

wrong-answer reports go to tech review before clinical staff

Architecture

Frontend
Next.js
Tiptap editor
↓
Backend
Server actions
Answer model (Sonnet)
Verifier (Haiku)
Import wizard
↓
Data
PostgreSQL
Prisma
↓
External
Anthropic API
AWS SES
Cloudflare R2

By the numbers

Prompt caching keeps a full-manual assistant affordable

Input token cost, no caching100%
Input token cost, with caching~10%

Context

A healthcare team's operations manual is consulted daily and must stay accurate. In a clinical setting a confidently-wrong answer is a safety problem rather than a usability one, so a naive 'chatbot over our docs' was unacceptable.

What I did

  • Built grounded Q&A that loads the full manual as cached context and answers with section citations, removing retrieval-miss risk for a bounded corpus.
  • Added an independent second model that verifies every answer is supported verbatim, returning a boolean gate before anything ships.
  • Designed document import so the LLM emits only structure and anchors; section bodies are spliced from the source, removing rewrite/hallucination risk.
  • Wrapped high-stakes generated artifacts in a seven-step clinical release audit (banned-term + special-character scanners, mandatory fact-verification, hard-stop rule).

Outcome

  • Staff get answers they can rely on, with wrong-answer reports triaged technically before ever reaching clinical decision-makers.
  • Clinical words in the manual are always the human-authored source words; the model never authors clinical text.
  • Prompt caching keeps the assistant cheap enough to leave running all day.

Under the hood

Grounding by a second model, not by vibes

  • The answer model responds only from the approved manual with citations; a separate, cheaper model then verifies each claim is supported verbatim and returns a boolean.
  • Any answer that comes back unsupported is flagged before it is shown, turning 'a chatbot over our docs' into an assistant clinical staff can rely on.

Verbatim-splice import

  • To structure an existing Word/PDF manual, the LLM emits only section titles and exact start-anchor quotes, and the bodies are spliced from the source between anchors.
  • The model never authors clinical text, and unmatched sections surface as explicit gaps for a human to resolve.

Clinical release audit

  • High-stakes generated artifacts pass a seven-step audit: banned-term and special-character scanners, cross-document consistency, and mandatory primary-literature fact checks.
  • The governing rule is simple: a single unresolved clinical-fact issue is a hard stop.

Stack

Next.jsPostgreSQLPrismaAnthropic (Sonnet + Haiku)TiptapPlaywright to PDF