Rebuilding a $1.5M medical AI pipeline with six LangGraph agents
MaryamMiradi · x · 2026-10-29
Dr. Maryam Miradi rebuilt a medical AI agent pipeline from a paper using Python and LangGraph, claiming it can replace systems that cost hospitals $1.5M/year yet fail in deployment.
The core problems: clinicians don't trust black-box AI, and data scientists waste 80% of their time cleaning messy data.
The six-agent design (see image):
- File Reader: auto-reads CSV/Excel/JSON/ZIP and routes files
- Privacy Guard: detects and masks names, emails, patient IDs, SSNs; redacts text in medical images
- Data Reader: parses spreadsheets and infers column meanings
- Model–Data Feature Matcher: assigns unique IDs per column and matches each dataset to the best AI model
- Data Preparer: labels columns, cleans large datasets (>50MB), matches image modalities
- Predictor + Explainer: LLM for small data, deep learning for big; highlights key regions in medical images so doctors see the "why", not just answers
She also provides a step-by-step build flow for each agent — a useful reference skeleton for deploying medical AI.
More from coding & agent
- GitHub's zeeg: worktrees are the poor man's sandbox for coding agents — rsasaki0109 · 2026-10-29
- Building with coding agents and middle management are wildly different skills — corbtt · 2026-10-29
- Replit's ViBench 2.0: top models pass only 49-65% of long-horizon builds — amasad · 2026-10-29
- Interactive 3D biology lesson built with Opus 5.5 and Three.js follows one breath — techartist_ · 2026-10-29
- Replit's ViBench 2.0: best models pass only 49–65% of builds, GPT-6.1 Sol leads — amasad · 2026-10-29
- Dev calls for task-level undo for personal agents, like database transactions — gethackteam · 2026-10-29