Building HelloMed: Lessons From a Hospital Management System
Designing relational schemas, integrating LLMs, and the small details that make healthcare software actually usable.
Why hospital software is hard
Software for healthcare has constraints most apps don't: audit trails, role separation, and zero tolerance for silent data corruption. Building HelloMed taught me to think in workflows, not screens.
The relational schema
The trick is to model the encounter, not the patient. Each visit, prescription, and lab order belongs to a specific encounter, which keeps history clean and audit logs straightforward.
patients ──< encounters ──< prescriptions
│
└──< lab_orders ──< lab_results
AI-assisted recommendations — carefully
I integrated an Ollama-powered recommendation engine that suggests differential diagnoses based on symptoms. Important constraints:
- The AI suggests, never decides.
- Every suggestion is logged with the prompt and model version.
- Doctors can override with one click, and the override is part of the audit trail.
What I'd do differently
If I started again, I'd build the event log first. Healthcare software is unforgiving when you can't reconstruct what happened.
TL;DR
Schema first, audit log first, AI last. Then layer the UI on top.