Automating the boring parts

AI engineer · LLM systems & agents

Pune, IN ·

Medipal

Upload a lab report and get a plain-language reading of it, powered by a fine-tuned model.

Year
2024
Role
Hackathon team project
Type
COMSYS hackathon runner-up
Medipal chat interface explaining a patient's uploaded medical report

Overview

Medipal is a medical report analyser our team built at the COMSYS hackathon, where it placed runner-up. You upload a PDF lab report, and Medipal reads it, explains possible findings and their likely causes in plain language, and lets you ask follow-up questions in a chat.

How it works

  1. Upload. A React (Vite) frontend with Clerk for authentication accepts the PDF report.
  2. Parse. An Express backend extracts the report text with pdf-parse.
  3. Analyse. The text goes to a GPT-3.5 model fine-tuned on the MedMCQA medical question dataset. It is prompted to answer in a fixed format: possible indications, likely causes, and a recommendation to seek further examination.
  4. Ask follow-ups. The chat keeps the report as context for follow-up questions. History stays in the browser's local storage.

Engineering decisions

  • Fine-tuned, not just prompted. Training on MedMCQA gives the model medical vocabulary and a consistent, clinical tone.
  • A fixed response shape as a guardrail. Every analysis ends by recommending further examination. It's a small constraint, but it matters for anything health-related.

Built with

  • React
  • Vite
  • Node.js
  • Express
  • OpenAI fine-tuning
  • Clerk

Next project

MindVault

A private, AI-organized second brain. Save anything in two seconds, find it again by meaning.