abi2209mj/Ticker_RL_Environment

🎫 Customer Support Ticket Resolution — RL Environment

A real-world reinforcement learning environment where an agent resolves customer support tickets through multi-turn dialogue.

Built for the OpenEnv Hackathon — fully OpenEnv-compliant.

🌟 Key Features (Beyond Baseline)

1. 🧠 Dynamic LLM Ticket Generator

Every reset() calls Claude to generate a fresh, realistic ticket — different industries, emotions, and complexity. No two episodes are the same.

2. 😤 Sentiment Escalation Engine

Customer sentiment degrades dynamically in response to bad agent behavior:

• —Repeated questions → sentiment worsens
• —Wrong solution → sentiment worsens
• —Correct resolution → sentiment improves
• —Sentiment state: happy → neutral → frustrated → angry → demands_supervisor

3. ⏱️ SLA Step Budget Penalty

Each task has a step budget (Easy: 3, Medium: 5, Hard: 8). Steps over budget incur a -0.05 per-step penalty, rewarding efficiency.

4. 📊 Live Gradio Dashboard

Deployed on HuggingFace Spaces — interactive browser UI with:

• —Live conversation replay
• —Real-time reward breakdown table
• —Cumulative reward chart
• —Sentiment timeline chart
• —Session leaderboard

📐 Observation Space

⚡ Action Space

🎯 Reward Function

📦 Tasks

🚀 Setup & Usage

# Install
pip install -r requirements.txt

# Set API keys
export ANTHROPIC_API_KEY=your_key   # for LLM ticket generation
export HF_TOKEN=your_hf_token       # for baseline inference

# Run the dashboard
python ui/dashboard.py

# Run baseline inference
python scripts/baseline_inference.py

🐳 Docker

docker build -t support-rl-env .
docker run -p 7860:7860 -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY support-rl-env

🤗 HuggingFace Spaces

Deploy by pushing to a HuggingFace Space with gradio SDK. Tag the space with openenv for discoverability.

📊 Baseline Performance