RxExplainer

A mobile-first web app where Spanish-speaking patients photograph a prescription label and receive an immediate plain-language explanation in both text and natural-sounding Spanish audio.

Live demo: https://rx-explainer.vercel.app

CS 153 final project — Stanford, Spring 2026

---

What it does

1. Patient photographs their prescription label
2. Claude Vision OCR extracts drug name, dosage, frequency, and warnings
3. A few-shot prompted model generates a warm, plain Latin-American Spanish explanation (6th-grade reading level)
4. ElevenLabs reads the explanation aloud
5. Every explanation is scored and stored in Supabase; a self-improvement loop rewrites low scorers and promotes high scorers into the example bank

Research question

Can a few-shot prompted model with a curated domain-specific dataset outperform zero-shot Claude on plain Latin-American Spanish medical explanations — and can the system self-improve over time using automated metrics?

---

Eval results

Supabase — live explanations table

Zero-shot Claude vs RxExplainer (few-shot + evolved example bank)

To isolate the contribution of the curated few-shot example bank, we ran a controlled zero-shot baseline: the same Claude model, the same 5 held-out labels, the same system prompt — but with no examples in context. RxExplainer's evolved bank outperforms zero-shot by +3.2 composite points, with the gain concentrated in accuracy (+6.6 points). This confirms that the curated examples are teaching the model a specific behavior — reliably surfacing drug name, dosage, and frequency in the output — that prompt engineering alone does not produce. Readability and tone are statistically equivalent across conditions, which is the expected result: dialect adherence is achievable through the system prompt, while field-level accuracy requires example-driven grounding.

Learning curve — composite score over 5 evolution cycles (n=5 held-out labels)

Composite score improves from 69.8 → 78.0 (peak at cycle 4) as the example bank grows through automated evolution. The gain is driven entirely by accuracy (+20 points over 5 cycles) — as high-scoring explanations are promoted into the few-shot pool, the model receives progressively stronger grounding for field inclusion. Tone holds at 100 across all cycles, validating that Latin American Spanish dialect adherence is stable once established. Note: n=5 held-out labels; results are directional.

---

Scoring system (fully automated, no LLM-as-judge)

Metric	Formula	Target
Readability	Fernández Huerta (Spanish-adapted Flesch-Kincaid)	80–100 = patient level
Accuracy	% of drug name / dosage / frequency found in explanation	Higher is better
Tone	Penalize Castilian Spanish words (vosotros, coger, etc.)	100 = clean
Composite	0.4 × readability + 0.4 × accuracy + 0.2 × tone	>80 = promoted to example bank

---

Tech stack

Layer	Tool
Frontend + API routes	Next.js
Vision OCR + explanation	Anthropic Claude (claude-sonnet-4.6)
Text-to-speech	ElevenLabs (eleven_multilingual_v2)
Database	Supabase (Postgres)
Hosting	Vercel

---

Running locally

npm install

Create a .env file:

ANTHROPIC_API_KEY=...
ELEVENLABS_API_KEY=...
SUPABASE_URL=...
SUPABASE_ANON_KEY=...

npm run dev

Open http://localhost:3000.

Run the zero-shot benchmark

npx tsx scripts/zero-shot-benchmark.ts

Run the learning-curve experiment

npx tsx scripts/learning-curve.ts --cycles 5