README.md

metadata

title: PharmaAgent
emoji: 💊
colorFrom: blue
colorTo: green
sdk: docker
pinned: false

🏥 PharmaAgent — Clinical Decision RL Environment

An OpenMV-compatible reinforcement learning environment where an agent acts as an AI clinical pharmacist, navigating a multi-step patient treatment pipeline powered by real DrugBank data.

What the Agent Does

Given a patient case (symptoms + existing medications), the agent must:

Step	Action	Reward
1	Diagnose — identify the condition from symptoms	0–0.30
2	Select drugs — build a safe treatment regimen	0.05–0.20 per drug
3	Check DDI — verify drug-drug interactions	0.05–0.30
4	Finalize — submit the complete regimen	0–0.30

Total possible reward per episode: 1.5 (normalized to 0–1)

Why This Is Different

• All scoring is grounded in real DrugBank data (19,842 drugs, 500k interaction pairs) — no hardcoded opinions
• Patient cases are generated dynamically from the database — the agent cannot memorize them
• The LLM is called only once per episode, at the end, purely to format a human-readable summary — it has zero influence on any reward
• Penalizes dangerous drug combinations using DrugBank's own severity classifications

API Endpoints

Method	Path	Description
POST	/reset	Start new episode
POST	/step?session_id=X	Take an action
GET	/state?session_id=X	Get session state
GET	/health	Health check
GET	/web	Browser UI for manual testing

Stack

• FastAPI + SQLite (DrugBank)
• Qwen 2.5 72B via HuggingFace Inference Router (agent)
• Groq Llama 3.3 70B (summary formatter only)