On-device medical AI · QVAC Hackathon
Works alone. Stronger together.
Pharos reads a medication label, normalizes it to its generic name, checks it against your saved shelf, and explains the risk in plain language — entirely on your phone. When a capable device is nearby, harder cases are delegated to a bigger model over a private peer-to-peer mesh.
⚕️ Educational information only — Pharos surfaces documented interactions to help you talk to a professional. It does not diagnose, prescribe, or recommend treatment.
One product, two tiers
The floor is offline. The ceiling is a mesh.
Solo tier — any one phone
Camera → on-device OCR → generic-name normalization → DDInter interaction lookup → an on-device MedPsy model explains the result. No account, no cloud. After a one-time model download it runs in airplane mode, provable with a zero-traffic network capture.
Mesh tier — stronger together
When a nearby anchor is present, the explanation is delegated over a Holepunch DHT to a larger MedPsy-4B running on that device — OCR and the interaction lookup stay on the phone. If the anchor is unreachable, it falls back to the on-device model.
Safe by construction
It retrieves facts — it doesn't invent them.
Retrieve, then explain
Interactions come from the DDInter 2.0 dataset, not the model. The model only explains a fact that was already retrieved and cited.
Abstains when unsure
If a drug can't be resolved to the dataset, Pharos says so rather than guessing. Unknown ≠ safe; no-match ≠ safe.
Your data stays yours
The medication shelf is stored in the device keystore. Nothing about your medications leaves your phone.
Verified, not asserted
What's actually been proven
- The grounded pipeline (resolve → lookup → abstain → audit) passes a 38-check deterministic suite.
- The real engine runs on Tether QVAC 0.12.2: live OCR → DDInter Major, plus abstain and no-fabrication, validated on real inference.
- The mesh tier is proven across two physical machines: a phone-class consumer delegated to a 4B anchor over the DHT and received the answer from the anchor.
- No remote API calls for inference — all OCR and MedPsy run locally.