— System Simulation
Hospital Simulation
LLM-powered hospital staff diagnose synthetic patients through a step-by-step workflow, visualized as characters moving through a 2D pixel-art hospital.
PydanticAI Python Phaser.js State Machines Canvas Optimization Multi-agent Systems CSS Modules AI Document Parsing
~ when Oct 2025 – Dec 2025
~ status Archived
~ team collaborative
~ kind Internship Project
A multi-agent AI simulation where LLM-powered hospital staff diagnose synthetic patients through a step-by-step workflow, visualized as animated characters moving through a 2D pixel-art hospital. Built during an internship at LifeAtlas, in a two-person core team with Andreas.
Agent Pipeline
- Nurse triages incoming patients and routes to the Doctor
- Doctor examines, orders tests (routed to Lab), and diagnoses - informed by vector search over past cases and learned principles via ChromaDB
- Lab Tech processes ordered tests and returns results
- Reflection Agent reviews completed cases, identifies diagnostic errors, and stores learning principles for future retrieval
- Pipeline orchestrated with guardrails: max test cycles, duplicate test detection, step limits
Learning & Memory
- ChromaDB vector search lets agents query past successful cases and past mistakes to inform current diagnoses
- Experiences tracked with retrieval counts and accuracy metrics
- Low-performing diagnostic principles can be automatically deprecated
Visualization
- Phaser.js 2D engine with a Tiled-authored 32x32px tileset hospital map
- Custom A pathfinding* on the tile grid (Manhattan distance heuristic)
- Staff NPC waypoint system with idle actions and speed multipliers
- Patient queue managing active/waiting/completed states
Document Extraction
Patients can be created from uploaded documents - supports PDF (with embedded image OCR via Tesseract), DOCX, RTF, ODT, and spreadsheets.
About this chart
Each axis is a functional pillar; the orange area is where my focus went, and the purple how much of that work was AI-augmented. That AI layer is where tools sped up implementation - architecture, code review, and the quality bar stay mine. I treat AI as a precision tool with strict conventions, not auto-pilot.