PARAKH (परख): Food Label Reader for India

PARAKH is a Python web app that reads Indian packaged food labels from photos and gives an honest health breakdown in under 30 seconds. You upload one to three photos of a package and choose English or Hinglish. It uses Gemma 4 (26B) through the Google AI Studio API.

What you get:

Every ingredient is rated as good, moderate or avoid.
Sugar, salt and fat get traffic-light grades.
Expiry is calculated with a countdown in days.
The veg or non-veg mark is detected.
The maker’s claims are checked against the label text.
You can ask questions like “Is this safe during pregnancy?”
You can download a full report.

How it works: Gemma 4 reads the photo and pulls out the text, nutrition table and dates. Code then checks that every ingredient it reported actually appears on the label, and removes any it can’t find. Python, not the AI, calculates the expiry dates and nutrition grades. A curated list of about 300 additives is used for ratings, and the AI only fills the gaps.

Why it can be trusted:

Hallucinated ingredients are dropped.
Unreadable text is marked [unclear].
Claims like “100% safe” are stripped out.
Medicine labels are refused.
Answers are marked as coming from the label or from general knowledge.
Photos are processed in memory only, with location data removed, and nothing is stored.

It has 72 passing tests, and it is built with Gradio, Pillow, RapidFuzz and Pydantic.