Genium: verify medicine and food before you use it.

Fake and mislabelled medicines, herbals and packaged foods are a daily risk in Nigeria. NAFDAC’s Greenbook registry exists, but it’s a manual lookup almost nobody uses, and registration numbers get stolen: a counterfeiter can print a valid number from a cheap paracetamol on a fake Augmentin. “The number exists” doesn’t mean “this pack is real.”

What it does. Genium lets anyone snap a photo of a pack and get a plain-language verdict in English, Yoruba, Hausa, Igbo or Pidgin, with a clear next step.

1. Read: Gemini-3.5-flash Vision extracts the NAFDAC number, product name, manufacturer, batch and dates, with a confidence score per field.
2. Verify: deterministic checks run against a Greenbook snapshot. Does the number exist? Does it belong to this product and manufacturer? Are the dates valid? Does the carton batch match the blister batch?
3. Crowd intelligence: reports are keyed by product and batch, so one suspicious sighting raises the risk for everyone holding that batch.
4. Explain: Gemma 4 turns the findings into a faithful, safe explanation in the user’s language.
5. Learn: Gemini-3.5-flash parses free-text reports in local languages into structured records, and generates localized training rows into a live .jsonl dataset.

A verdict that’s honest about its limits. Genium never says “this is fake.” It returns Green (registry and batch clear), Amber (label matches, but the batch has crowd-reported issues or unusual scan activity) or Red (number missing, mismatched or packaging contradictory). When it can’t tell, it says so and sends the user to a pharmacist or NAFDAC. It gives no medical advice.

Catching clones. A phone camera isn’t a spectrometer, so we layer defenses: registry cross-checks catch stolen numbers, two-surface photos (carton plus blister) catch mismatched packaging, batch velocity and geo-collision anomalies catch one real batch number copied across thousands of fakes, and a crowd kill-switch catches even perfect visual clones. A counterfeiter can clone ink on cardboard, but they can’t hide a bad batch from a connected crowd.

Tech and sponsors.

• Gemma 4 and Gemini-3.5-flash: vision extraction, multilingual explainer, report parser and synthetic data expander.
• Adaption (Adaptive Data workflow): every scan and report generates structured, localized training rows, ready for AutoScientist fine-tuning. We mirrored the workflow with Gemma 4 and Gemini-3.5-flash in a compatible format.
• DigitalOcean: hosts the API, registry snapshot and dataset storage.

Why it scales. Every scan and report improves batch intelligence at near-zero marginal cost. The same pipeline extends to drugs, food, cosmetics and herbals, new languages need only new seed data, and the live dataset becomes a fine-tuning asset for small open-weight models in Nigerian languages.

Demo. Scan a genuine pack for a green pass in about 2 seconds. Then scan a mocked “Augmentin” carrying Emzor Paracetamol’s NAFDAC number: red mismatch, read aloud in Yoruba. Finally, submit a one-sentence Yoruba or Pidgin report on that batch and watch the hotspot dashboard spike to High Risk while the dataset drawer fills with new multilingual training rows, live.

Honest limits. Photos verify labels and registry consistency, not chemical content. The demo registry is a curated snapshot of about 50 products, and production would sync with NAFDAC data. Crowd signals need volume, which we handle with thresholds and verified-pharmacist weighting.