MedCheck — Vision-Assisted Medication Label Verification
An OCR-first medication-label verifier built for HackRU 2026.
The problem
A caregiver needs to answer a narrow question: does this bottle match the tracked medication, and is a dose actually due? Reading a label is not the same as making a medication decision, and a model should not silently bridge that gap.
MedCheck is an OCR-first label-verification prototype built at HackRU 2026. It turns a photo into reviewable evidence, compares it with a tracked prescription, and requires a separate human confirmation before any dose is recorded.
Safety invariants
The architecture is shaped by explicit constraints:
- A scan is evidence about a bottle, not proof of ingestion.
- Inference never authorizes a dose.
- The backend owns prescriptions, schedules, inventory, and dose events.
- Structured fields preserve confidence and source metadata.
- Missing or ambiguous name/strength evidence produces
uncertainor a rescan, not a guess. - A scan never fabricates drug data; raw source text is retained.
These are product requirements, not claims that the prototype has been clinically validated.
How it works
photo(s)
→ frame-quality checks + YOLO11n bottle detection
→ RapidOCR label extraction
→ structured fields with confidence and source
→ multi-frame fusion for long/curved labels
→ optional Qwen3-VL escalation when identity is weak
→ candidate match against tracked prescription
→ review / mismatch / uncertain / due-now state
→ explicit confirmation before dose history
The inference service reports what it can read; the Flask/TimescaleDB backend owns medication authority. A React/Vite dashboard supports scan, verify, review, confirm, history, supply, and adherence views. A separate SwiftUI preview is described in the README, but the audited repository's provenance record says no iOS app is present there; treat the iOS surface as a preview, not a shipped client.
Contribution and provenance
I originated and directed the project: I set the OCR-first architecture, the safety invariants, and the integration plan, and coordinated the work across the dashboard, backend, and inference services. Teammates and agent-assisted implementation contributed parts of the codebase; the repository's docs/PROJECT_PROVENANCE.md records the authorship and integration history. This page does not attribute every file to me.
Evidence and limits
- The repository's release manifest reports 523 inference tests and 81 backend tests at commit
57eb80b. Those numbers were documented by the project; they were not re-executed for this portfolio build. - A UI-only mock-data mode exists for design review. It is not the live verification pipeline.
- The full stack is a local development environment, not a hosted medication service.
- No clinical validation, medication-correctness accuracy, or patient-safety claim is made.
- OCR and VLM output can be wrong; confidence, provenance, and human review are the mitigation, not a guarantee.