Architecture & Methodology
How MedAssist turns unstructured PDFs into clinical intelligence.
This project was built for a Design Engineering capstone to solve a specific clinical UX problem: the gap between receiving lab results and understanding them. Here is how the pipeline works.
OCR & Text Extraction
The uploaded PDF is routed to a secure, ephemeral serverless function. We use advanced OCR and layout parsing to retain the hierarchical structure of the lab report, identifying panels (e.g., CBC, CMP) and extracting tabular rows (Biomarker, Value, Unit, Reference Range).
LLM Pipeline & Normalization
Extracted text is processed by a medical-grade LLM pipeline. The pipeline normalizes biomarker names to standard schemas, calculates percentage deltas against the patient's historical baseline, and flags values that fall outside the clinical reference ranges provided by the specific lab.
Insight Generation
Based on the normalized data, the system generates plain-English insights. It identifies key clinical correlations (e.g., high LDL + high Triglycerides) and drafts actionable questions the patient can bring to their physician.
Secure Persistence
Data is stored in a PostgreSQL database protected by Row-Level Security (RLS). Every report is cryptographically bound to the user's session, ensuring no cross-tenant data leakage.
Built with Security First
Our architecture guarantees that your data is never used to train foundation models, and is protected by industry-standard encryption.
Read our Security Principles