MedLens

Inspiration

Medicine labels contain authoritative information, but they are often dense, repetitive, and difficult to scan. I wanted to make this information easier to navigate without replacing the official source or allowing AI to invent medical claims.

What it does

MedLens is a consumer health-literacy app for U.S. patients and caregivers. A user searches for a brand or generic medicine, selects the exact product type, route, dosage form, and FDA label, and receives a concise five-part dashboard (Uses; Warnings; Do not use; Side effects; Interactions). Every visual fact includes a numbered citation. Selecting it opens the exact, unchanged FDA passage supporting that fact. Users can also open the complete FDA/openFDA record and corresponding NIH DailyMed label. MedLens does not provide diagnosis, dosing, personalized recommendations, or pill identification.

How we built it

I built MedLens in a continuous Codex workflow using GPT-5.6. Codex accelerated the React and TypeScript implementation, FDA integrations, API architecture, responsive interface, accessibility, automated tests, debugging, and deployment. It was especially useful for turning safety requirements into testable application rules and iterating quickly against real FDA label edge cases. At runtime, GPT-5.6 performs two separate tasks:

1. Structured extraction of concise facts from the selected official label.
2. Independent review of whether each simplification is directly supported by its quoted evidence. Structured Outputs constrain the response format, while application code performs deterministic checks. A fact is displayed only when its quotation appears continuously in the official FDA text, its category is allowed for that source field, and it passes the independent review. Unsupported claims are withheld rather than filled from model memory.

Challenges we ran into

FDA labels vary significantly between prescription and over-the-counter products. Information can appear in different fields, some products have no dedicated adverse-reactions section, and several labels repeat the same warning in multiple places. The hardest challenge was preventing one source statement from becoming several apparently independent visual facts. I used Codex to develop evidence-overlap checks, warning-topic fallback deduplication, focused extraction passes, and regression tests for these cases. Another challenge was balancing readability with traceability. The interface needed minimal text, but users still needed a clear path from every visual card to the full unchanged evidence.

Accomplishments that we're proud of

MedLens makes the evidence visible instead of hiding it behind a generic “AI-powered” claim. The visual dashboard and FDA source map use the same five categories, colors, counts, and citation numbers, making every simplification auditable. The completed project includes responsive layouts, keyboard-accessible interactions, loading and error feedback, bounded retries, strict source validation, and automated tests covering official-source failures and medical-label edge cases.

What we learned

Structured output does not guarantee factual correctness. Reliable AI applications need validation outside the model, visible provenance, carefully limited scope, and honest failure states. I also learned that medicine-name search alone is insufficient: products sharing an ingredient may have different routes, formulations, manufacturers, and official labels.

What's next for MedLens

Future work could add deterministic PDF export, Spanish interface text, RxNorm-assisted name disambiguation, and non-persistent comparison of two formulations—while preserving the same source-first safety model.

Built With

• codex
• css
• gpt-5.6
• nih-dailymed
• node.js
• openai-api
• openfda-api
• responses-api
• structured-outputs
• typescript
• vinext
• vite