Skip to content

biopanel.io — AI-Powered Health Analytics Platform

Project Summary

Project: biopanel.io
Type: Personal product — built and owned end-to-end
Stack: Python · FastAPI · Celery · Redis · PostgreSQL · Supabase · React 18 · TypeScript · Docker · LLMs

What it does: Extracts structured biomarker data from blood-test PDFs across any laboratory, language, or document layout — with zero hardcoded parsing rules. Users upload a lab report, the system normalizes all values, flags out-of-range markers, and renders an interactive health dashboard.

The Problem

Blood test results come in hundreds of different PDF formats — different labs, different languages, different layouts, different reference ranges. Traditional extraction approaches require lab-specific parsing rules that break the moment a format changes.

The challenge: build a system that extracts biomarkers reliably from any PDF, without knowing the layout in advance, at production-grade reliability.

Architecture & Approach

I designed an asynchronous extraction pipeline that treats each PDF as a structured extraction problem for a language model — not a pattern-matching problem.

PDF Upload → Celery Task → PDF Parser → LLM Extraction → Normalization → PostgreSQL
                                              ↓
                                    Multi-provider fallback
                                    GPT-4o → Claude Sonnet

Key architectural decisions:

  • Typed DAG via Celery + Redis: each stage is an explicit async task with typed inputs/outputs — no silent failures
  • Multi-provider LLM fallback: GPT-4o primary, Claude Sonnet as fallback — resilient to provider outages
  • Streaming via Server-Sent Events: extraction results stream to the frontend as they arrive — no polling
  • pgvector for biomarker search: normalized biomarkers stored with vector embeddings for semantic search
  • Zero hardcoded rules: the LLM handles layout variability — the pipeline handles structure and validation

Tech Stack

Backend

  • FastAPI — async REST API with full OpenAPI docs
  • Celery + Redis — async task queue with retry logic and typed DAG
  • PostgreSQL + pgvector — persistence and vector search
  • Supabase — auth and storage layer
  • Alembic — database migrations
  • Docker Compose — local development and deployment

Frontend

  • React 18 + TypeScript — full type safety end-to-end
  • Custom chart components for biomarker visualization
  • Responsive mobile-first layout
  • Real-time SSE for extraction progress

AI / LLM

  • OpenAI GPT-4o — primary extraction model
  • Anthropic Claude Sonnet — fallback model
  • Custom prompt engineering for structured biomarker extraction
  • LangChain for chain orchestration

Results

  • Extracts biomarkers from PDFs across multiple labs, languages, and layouts with no layout-specific configuration
  • End-to-end extraction completes in under 30 seconds for a typical lab report
  • Multi-provider fallback ensures availability when individual providers have outages
  • Deployed as a public product at biopanel.io

What this demonstrates

This project exists to prove a specific capability: I can take an idea from zero to a production AI product — architecture, backend, frontend, deployment, and all. It's the most complete demonstration of my full-stack + AI engineering depth.

For data engineering clients: the same pipeline design patterns (async DAGs, typed stages, observability, fallback logic) are how I approach enterprise data workflows.

  • Interested in AI-powered data systems?

    Let's talk about what this kind of architecture could look like for your use case.

    Book Intro Call