PROJECT_SUMMARY.md

Project Summary: MedTriage OpenEnv

Meta PyTorch OpenEnv Hackathon Submission

🏥 Project Overview

MedTriage is a real-world simulation environment designed for training and evaluating AI agents in clinical decision-making. Unlike "toy" environments, it models a genuine human task: categorizing patient symptoms into appropriate care levels (Self-Care, Clinic, Urgent Care, or Emergency).

🛠️ What We Built

We have developed a complete, spec-compliant OpenEnv environment with the following components:

1. Core Environment (server/triage_environment.py)

• Action Space: A triage_patient tool where agents assign a triage level (0-3) and provide clinical reasoning.
• Observation Space: Detailed patient profiles including demographics, unstructured symptom text, vitals (BP, Temp, HR, SpO2), and medical history.
• Reward Function: A medical-safety-first scoring system (0.0 - 1.0) that rewards accuracy and penalizes dangerous under-triage (e.g., missing an emergency).

2. Mandatory Tasks

Implemented three specific scenarios with automated graders:

• TASK_EASY: Seasonal Allergies (Self-Care)
• TASK_MEDIUM: Possible Appendicitis (Urgent Care)
• TASK_HARD: Atypical Myocardial Infarction in elderly patient (Emergency)

3. API Infrastructure (server/app.py)

• Implemented standard OpenEnv endpoints: /reset, /step, /state.
• Added required Hackathon endpoints:
  • /tasks: Dynamic list of available scenarios and schemas.
  • /grader: Real-time performance scoring.
  • /baseline: Automated trigger for the inference script.

4. Baseline Inference (inference.py)

• Created a reproducible baseline script that leverages the Hackathon's LiteLLM proxy to evaluate the environment. It ensures the environment is "solveable" using a real LLM and serves as a benchmark for API-based agent interactions.

🚀 Technical Improvements & Fixes

To ensure the project meets the highest submission standards, we performed the following:

• Spec Compliance: Added missing __init__.py files and standardized the directory structure to pass openenv validate.
• Dependency Management: Updated pyproject.toml and server/requirements.txt to include openenv-core>=0.2.0 and generated a uv.lock file.
• Containerization: Optimized the Dockerfile to support root-level builds, enabled health checks, and added editable installation (pip install -e .) to resolve local module imports.
• Deployment Automation: Configured README.md with Hugging Face Space metadata (app_port: 8002, sdk: docker) for seamless hosting.

🌐 Deployment Links

• GitHub Repository: https://github.com/ash399/med-triage-openenv
• Hugging Face Space: https://huggingface.co/spaces/ashdev/med-triage-openenv

---

Documented on: Monday, March 30, 2026