MedCheckRx

The Problem

MedCheckRx needed a way to verify pill identity during pharmacy fills without slowing down an already-efficient workflow. Generic computer vision models trained on public datasets failed to meet the accuracy bar the pharmacy required — pill appearance varies by manufacturer lot, lighting conditions, and tray orientation in ways that confound off-the-shelf models. The solution had to fit inside their existing verification protocol, not replace it.

The Approach

We built a CV pipeline trained specifically on MedCheckRx's own verification protocol. Cameras capture the fill tray at the exact point in the workflow where a pharmacist would typically review it. The model verifies pill identity against the expected medication and flags any exceptions for human review rather than attempting to auto-resolve them. By encoding the pharmacy's specific domain knowledge into the training data and model logic, we achieved accuracy that generic models cannot reach on this task.

• —Custom CV pipeline trained on pharmacy-specific verification workflows
• —Workflow-embedded logic that matches real fill-and-verify steps
• —Automated exception flagging for human review on ambiguous cases
• —Continuous accuracy improvement loop using production data
• —Production-grade deployment with monitoring and alerting

The Outcome

MedCheckRx is an active production system. The pharmacy runs it on every fill, and the system catches errors that generic models miss — because it was trained on their specific workflow, not on a public pill dataset. We continue as their engineering retainer partner, improving accuracy as new pill variants are encountered and the model sees more production data.