A role-based telehealth and prescription-management platform where doctors run Zoom consultations, generate digital prescriptions whose contents are auto-extracted from photos by GPT-4 Vision, and maintain per-condition patient timelines that drive automated follow-up reminders.
PharmaConnect targets the slice between 'WhatsApp + paper' and a full EMR: a digital prescription is generated alongside the consultation, GPT-4o-mini Vision extracts its contents from an uploaded photo, and every action (visit, lab, med change, follow-up) is pinned to a per-condition timeline. Doctors run Zoom calls; cron jobs fire follow-up reminders; four roles (admin, doctor, patient, pharmacist) operate through role-segmented dashboards.
Solo, end-to-end portfolio build across two repositories (a NestJS API + a Next.js App-Router client) and four roles (admin, doctor, patient, pharmacist). Sole owner of system design, data model, REST API, AI integration, frontend, auth, and DevOps — built to demonstrate a non-trivial healthcare workflow with third-party integrations (Zoom Server-to-Server OAuth, OpenAI GPT-4o-mini Vision, Google OAuth).
Paper prescriptions are lost, illegible, and have no audit trail: a misplaced slip means the dose history is gone, a misread drug name can't be disproven, and nothing ties a condition (e.g. Type 2 Diabetes) to the prescriptions, follow-ups, labs, and appointments issued for it. Small clinics can't justify a heavyweight EMR for this.
ChallengeGPT-4 Vision will happily invent medications if the source image is blurry, and a naive 'replace the patient summary' workflow risks deleting a real prior-medication history on every upload.
FixA two-call pipeline: (1) a Vision call extracts only what's visible in the image as structured fields; (2) a second text call merges that into the existing summary under a strict prompt contract that preserves historical medications and flags conflicts rather than overwriting — the model's role is additive, never destructive.
ChallengeZoom's Web SDK init can fail on slow networks or behind corporate proxies, and a non-joinable consultation kills trust in a healthcare app on the first call.
FixWrapped SDK init in try/catch and fell back to a static Google Meet link so the consultation still happens; the failure is surfaced to telemetry, and the backend Meeting table also stores join/start URLs so the doctor can share the link out-of-band.