Ali Raza
01AI · Health · Full-Stack · 2026

MediCare AI

AI-powered medical document intelligence — from scanned paperwork to structured, understandable information.

Status
Hackathon prototype
Year
2026
Role
Design, architecture, and full-stack implementation.
GitHub ↗Live demo ↗

01Context

Medical documents are often difficult to understand, poorly structured, and trapped inside images or scanned paperwork. MediCare AI converts them into structured, understandable digital information through an end-to-end pipeline.

The platform combines PaddleOCR for medical-document text extraction with Groq-powered AI for interpretation and patient-friendly explanations, while providing structured patient management and role-based portal access for individuals and healthcare organizations.

Built for the Alibaba Cloud AI Hackathon Pakistan 2026 as a prototype — explicitly not a substitute for professional medical advice, diagnosis, or treatment.

02Problem

Prescriptions, lab reports, and clinical records frequently arrive as images or scans — information that cannot be searched, stored structurally, or easily understood by patients.

The same document serves two audiences: an individual trying to understand their own medical information, and a healthcare provider who needs authorized access to patient records. Generic AI chatbots with a medical theme solve neither.

03Approach

Build a pipeline instead of a chatbot: real medical document → OCR → structured information → AI interpretation → patient/healthcare workflow.

Keep the OCR layer separate from the AI layer so extracted text can be processed independently and inspected before AI interpretation.

Frame AI strictly as assistance: source document → extraction → interpretation → human review. AI output is never treated as an unquestionable medical authority.

Separate Frontend, Backend/API, OCR + AI processing, and Data storage into distinct layers so compute-heavy OCR work and application traffic can be handled independently.

Enforce authentication, authorization, input validation, and server-side access control so patient information is only visible to authorized roles.

04Architecture

Document processing pipeline

  1. Medical Document

    Prescription, lab report, form, or scan

  2. Image Upload

  3. PaddleOCR

    Text extraction — independent of AI

  4. Text Extraction

  5. Structured Medical Information

  6. Groq AI

    Interpretation layer

  7. AI-Powered Interpretation

    Assistance, never diagnosis

  8. Patient / Healthcare Portal

    Role-aware access

05Stack

  • Next.js
  • React
  • TypeScript
  • Python
  • PaddleOCR
  • Groq AI
  • Supabase
  • JWT Auth
  • Vercel

06Features

  • 01Medical document OCR — extract text from prescriptions, laboratory reports, medical forms, and scanned documents.
  • 02AI-powered interpretation — medication information, dosage when present in the source document, terminology explanations, and patient-friendly summaries.
  • 03Individual / patient portal — upload documents, run OCR, view interpreted information, and manage medical records.
  • 04Hospital portal — authorized access to patient documents, OCR results, and AI-assisted information.
  • 05Separated OCR and AI layers so extraction can be inspected before interpretation.
  • 06Testing coverage around authentication, API behavior, OCR processing, and patient/hospital workflows.

07Engineering decisions

  1. D01

    Pipeline over chatbot

    Differentiation comes from connecting document intelligence, healthcare records, AI interpretation, and patient/provider workflows — not from providing another chat interface.

  2. D02

    Two independent processing layers

    OCR is kept separate from AI interpretation so extracted text can be verified on its own before any model sees it, and so the layers can scale independently.

  3. D03

    AI framed as assistance

    Every AI output path ends in human review. The README carries an explicit disclaimer: prototype only, no diagnosis, verify against the original document.

  4. D04

    Server-side enforcement

    Access control is enforced on the server, secrets stay out of source control, and errors are written so they do not leak sensitive information.

08Challenges

  1. C01

    OCR quality varies with real documents

    Document quality, formatting, handwriting, image resolution, and layout all affect extraction — which is why the pipeline inspects OCR output before AI interpretation instead of trusting it blindly.

  2. C02

    Healthcare data demands discipline

    Authentication, authorization, input validation, API validation, and secure secret handling are treated as requirements, not polish.

  3. C03

    Hackathon scope vs. honesty

    Documentation deliberately warns against claiming services that are not actually in the deployed architecture — the repository must reflect the real implementation.

09Result

  • Working prototype built and deployed for the Alibaba Cloud AI Hackathon Pakistan 2026 (medicare-ai-sigma.vercel.app).
  • End-to-end pipeline implemented: upload → PaddleOCR extraction → Groq interpretation → role-aware patient and hospital portals.
  • Repository includes e2e OCR pipeline tests and benchmark scripts alongside application code.
  • Explicit prototype disclaimer maintained — no medical diagnosis capability is claimed.

10What I learned

  • “Separating extraction from interpretation makes an AI system inspectable instead of magical.”
  • “In a healthcare-shaped product, what you refuse to claim matters as much as what you build.”
  • “A layered architecture (frontend / API / processing / storage) keeps a hackathon project honest about what is actually running.”