Teach Engine

Inspiration

Patients, caregivers, and even front-line staff keep getting healthcare documents they don’t fully understand — not just discharge notes, but EOBs, prior-auth letters, billing statements, prescriptions, and lab reports. Generic AI chatbots can explain these, but they don’t verify that the person actually understood. We wanted to turn the clinical “teach-back” method into an AI experience that works across clinical, pharmacy, and insurance documents, is installable (PWA), and doesn’t store PHI.

What it does

• Detects the type of document (prescription, EOB, prior auth, discharge, lab) and falls back to General if it’s something else.
• Simplifies the text to 6th–8th grade level using Gemini with Structured Output.
• Creates a 3–5 question quiz tied to the exact content in the document.
• If the user gets something wrong, it reteaches just that part and lets them retry.
• Shows a context-specific panel (e.g. EOB math, prescription dose/timing, discharge follow-ups).
• Has three modes: Teach-Back, Chat Helper, and Live Q&A (voice).
• Runs as a PWA so it can be installed and show the tutorial offline.
• Tracks only local, anonymous metrics (unknown docs, user overrides, mastery count).

How we built it

• Google AI Studio (Gemini) for classification + generation with Structured Output.
• React + TypeScript + Vite for a fast, component-based UI.
• PWA using manifest + service worker (vite-plugin-pwa) for installability.
• LocalStorage for privacy-friendly telemetry (no PHI stored).
• Accessibility-first layout: help modal, disclaimer modal, keyboard-friendly controls.

Challenges we ran into

• Preventing the model from forcing a wrong category when the document is out of scope → we added an explicit unknown path and a user override.
• Keeping the JSON strict so the UI can render the right domain card every time.
• Balancing PWA caching with not caching potentially sensitive responses.
• Designing one flow that works for clinical docs and payer/insurance docs.

Accomplishments that we're proud of

• One app now handles clinical, pharmacy, and insurance documents in a single UI.
• We don’t just explain — we measure comprehension (attempts + time-to-mastery).
• Voice/Live mode makes it usable for low-literacy or vision-limited users.
• Fully browser-based and deployable from Google AI Studio → good for demos, pilots, hackathons.
• We can show when the model was uncertain (unknown + override counters).

What we learned

• The real differentiator over generic chatbots is verification (teach-back), not simplification.
• A tiny “we’re not sure” banner + user override prevents most misclassifications.
• Structured Output is the key to adding more health document types later.
• PWA + AI Studio is enough to ship something real without standing up a backend.

What's next for Teach Engine

• Add more document families (imaging, consent, therapy homework, community-health flyers).
• Export to FHIR-safe summaries for EHR/payer portals.
• Run a small usability study and publish it as a perspective/short paper.
• Add a provider/admin view to surface common patient misunderstandings by document type.

Built With

• accessibility
• ai/ml
• cloud-run
• css3
• gemini-2.5-flash
• gemini-live-api
• google-ai-studio
• healthcare
• html5
• javascript
• localstorage
• pwa
• react
• tailwind-css
• typescript
• vite
• vite-plugin-pwa
• voice-ai
• workbox