PharmaGuide

The problem

Supplement labels don't say whether a form is bioavailable, whether an ingredient was recalled, or whether the evidence holds up. And a general-purpose chatbot is the wrong tool when someone asks about warfarin and ginkgo.

What we built

• A staged scoring pipeline

  Clean, enrich and score NIH label records. Ingredients are resolved to UMLS, PubChem, FDA UNII and RxNorm identifiers, and every stage is resumable with its own output contract.

• A six-pillar, 100-point score

  Formulation 20, Dose 20, Evidence 20, Transparency 15, Verification 15, Safety and hygiene 10. A banned substance always blocks a product, whatever its quality points.

• Deterministic before generative

  Emergency detection runs before the rate limiter. Known high-risk combinations resolve through 41 reviewed routes and 12 data-driven safety gates, and a validator checks every reply.

• Designed for failure

  Primary model, second tier, cross-vendor fallback and per-provider circuit breakers. When every provider fails, a deterministic safe reply is returned instead of an error.

• An app that shows, not recomputes

  The Flutter app displays pipeline scores as published, from an over-the-air catalog database, with a checksum-pinned interaction database and a separate local user store.

• Governed claims

  Clinical claims carry review dates and sources. A release gate in CI and on every deploy blocks a release when a claim is overdue.