NoteDoctor.AI
Prior Auth Engine.

A full-stack RAG application that generates prior authorization summaries from clinical inputs — with a split-panel UI, swappable layout, PDF export, authentication, and Stripe billing.

AI that reads the rules so physicians don't have to.

Prior authorization is one of healthcare's most expensive administrative burdens — 80M+ requests processed annually, with 40% of denials ultimately overturned on appeal. Physicians waste hours each week navigating payer-specific medical necessity guidelines. NoteDoctor.AI asked us to build the engine that changes that.

We designed and built a production RAG application from scratch. Clinicians enter a diagnosis, CPT codes, and patient history; the system retrieves the exact payer guidelines for that case, runs them through an OpenAI-powered generation layer via LangChain, and surfaces a structured prior authorization summary with cited medical necessity criteria — all in seconds.

Every request produces a structured readiness report: a request overview, clinical context, authorization and medical-necessity criteria, relevant codes, and a required-documentation checklist that marks what the note already covers, with a clear determination such as "More information needed". Reports can be saved, reopened, and exported as branded PDFs. The application also ships with full user authentication, Stripe subscription billing, light and dark themes, a fully responsive layout, and a swappable split-panel interface (Request/Report tabs with a Swap Layout toggle).

Building a RAG system accurate enough for clinical decision support — where retrieval errors have real patient and billing consequences — while delivering a polished, full-featured SaaS product.

Clinical accuracy and product quality at the same time.

Most RAG demos fall apart under real-world use. Healthcare is less forgiving than most — payer guidelines span hundreds of pages, update frequently, and vary by plan and state. A hallucinated medical necessity criterion or a missed CPT requirement isn't just a bad answer; it's a delayed or denied treatment. We built a multi-layer retrieval strategy (dense + sparse search, reranking before generation) and enforced strict source citation on every output. The system declines to answer rather than guess.

On the product side, the challenge was delivering a full SaaS application — auth, billing, responsive design, adaptive UI, and PDF generation — within the same eight-week timeline as the AI work. We used Next.js throughout, integrated Stripe for subscription management, and built the swappable split-panel layout to give clinicians flexibility in how they use the tool across different screen sizes and workflows.