Food For thoughts

Inspiration

Healthcare is one of the few domains where AI's potential is immense, yet the cost of getting it wrong is unacceptably high. I wanted to build something that augments clinicians rather than replacing them — a system that could handle the heavy lifting of image preprocessing, pathology detection, and report generation, but always with a human in the loop.

The name HomzDoctor reflects that vision: a doctor that feels like home — approachable, always available, but never overstepping its bounds. The 8‑stage workflow was designed to mirror real clinical pathways: from data ingestion → imaging pipeline → diagnostic inference → doctor review → patient communication → pharmacy → adherence tracking. Every step is built around the core constraint that only physician‑approved medications can move forward.

What I Learned

Building HomzDoctor was a crash course in several interconnected domains:

· Medical AI & Imaging — Working with DICOM, NIfTI, and medical imaging pipelines taught me the importance of robust preprocessing: volume reconstruction, N4 bias correction, normalisation, and spine segmentation. SimpleITK, pydicom, and nibabel became essential tools. · Human‑in‑the‑Loop (HITL) Systems — Designing the Doctor Review Layer forced me to think about UX for clinicians: approve, modify, or reject AI findings with clear controls. Building trust in AI means giving doctors the power to override anything. · Multi‑Agent Architectures — The platform uses 7 specialised agents (Orchestrator, Imaging, Diagnostic, Drug Knowledge, Pharmacy, Appointment, Patient Assistant). I learned how to coordinate them via RabbitMQ/Redis Streams, manage state, and handle failures gracefully. · Production‑Grade Full‑Stack — From FastAPI + PostgreSQL + Redis on the backend to React 18 + TypeScript + Tailwind on the frontend, and Docker + Kubernetes + Terraform for infrastructure — this project forced me to think about scalability, security, and deployment from day one. · Safety‑Critical Escalation — Implementing the Escalation Agent taught me how to define red‑flag triggers (chest pain, neurological deficit, severe allergic reactions, persistent high fever) and route them to immediate human intervention instead of continuing automated guidance.

Built With

• component-technology-backend-python-3.11+
• fastapi-database-postgresql
• github
• kubernetes
• medgemma-image-processing-simpleitk
• nibabel-frontend-react-18+
• pydicom
• redis-message-queue-rabbitmq-/-redis-streams-ai/ml-pytorch
• tailwind-css-devops-docker
• transformers
• typescript