README.md

metadata

title: Customer Support AI API
emoji: 🤖
colorFrom: blue
colorTo: green
sdk: docker
pinned: false

🎮 AI Customer Support Simulator

A strict, deterministic RL-style environment for training and evaluating AI agents on customer support tasks.

🏗️ Architecture

├── models.py              # Core data models (Episode, StepResult, Task)
├── environment.py         # Strict step enforcement + fail logic
├── inference.py           # Rule-based agent + CLI runner
├── app.py                 # Gradio UI
├── tasks/
│   ├── easy_task.py       # Delayed delivery complaint
│   ├── medium_task.py     # Billing double-charge
│   └── hard_task.py       # Angry VIP threatening legal action
├── graders/
│   └── base_grader.py     # BaseGrader + HardTaskGrader
└── tests/
    └── test_cases.py      # 4 mandatory test scenarios

🎯 Step Flow

Empathy → Collect Info → Investigate → Resolution

Each step must be completed correctly before advancing.

📊 Reward Formula

reward = base_score + step_bonus − penalty

Component	Value
Base (correct)	1.00
Base (wrong)	0.20
Step bonus	0.20
Max per step	1.20
Episode max	4.80

⚠️ Penalties

Trigger	Penalty
Wrong action in step	−0.30
Repeated response	−0.20
Early solution	−0.25
Angry customer ignored	−0.25
Generic / too-short response	−0.15
Wrong assumption	−0.20
Skipped step	−0.30

💥 Fail Conditions

• More than 3 wrong steps across episode
• Total reward below 1.5
• Agent stuck in loop: 3 consecutive failures at same step
• Single-step penalty ≥ 0.80

🚀 Local Setup

pip install -r requirements.txt
python inference.py --task all --agent rule   # CLI test
python app.py                                  # Gradio UI
python tests/test_cases.py                    # Run tests