Captured from the live deployment, not a mockup
01 · The problem
A patient cannot tell whether ninety-five rupees for a strip is fair, and confusing Celebrex with Celexa is a real pharmacy-counter failure. Built one flow from a photo of the strip to composition, fair price and a cheaper generic.
02 · How it works
- 01
Made abstention a product decision rather than a bug: when evidence is thin the app says so, and no price, dose or interaction is ever allowed to originate from the language model.
- 02
Covered the follow-on questions too, interactions across 68,639 pairs, duplicate therapy and look-alike name warnings, and documented where the source data is incomplete instead of hiding it.
- 03
Ask a general-purpose model to identify a medicine from a photo and it will invent a plausible answer. Built the pipeline so no price, brand, dose or interaction can originate from the model: those come from retrieval modules containing no model calls at all.
- 04
Identified medicines from a photo of the strip through vision OCR, returning composition, NPPA price check, Jan Aushadhi generic and a calibrated confidence, with an empty answer treated as a valid answer rather than a bug.
03 · What it cost, and what it returned
Checked interactions across 68,639 DDInter pairs with duplicate-therapy arithmetic, and surfaced look-alike sound-alike neighbours, the names close enough to be confused at a pharmacy counter.