Python Debug AI Agent

A production-grade end-to-end system for building a Domain-Specialized AI Agent that debugs Python errors using a fine-tuned Small Language Model (SLM).

Project Goal

Input: Python code + error traceback
Output: Structured debugging result:

{
  "error_type": "NameError",
  "explanation": "Variable referenced before assignment",
  "suggested_fix": "Initialize the variable before using it"
}

Architecture

1. Dataset Pipeline - Scrapes StackOverflow/GitHub for Python errors
2. SLM Fine-tuning - QLoRA on TinyLlama-1.1B
3. AI Agent - LangChain agent with AST tools + SLM
4. FastAPI Service - Production API endpoint
5. Evaluation Framework - Metrics & benchmarking

Project Structure

python-debug-ai-agent/
├── data_pipeline/          # Dataset creation
├── training/               # QLoRA fine-tuning
├── models/                 # Saved models
├── agent/                  # LangChain agent
├── evaluation/             # Metrics & eval
├── app/                    # FastAPI service
├── dataset/                # Generated data
├── requirements.txt
├── Dockerfile
└── README.md

Quick Start

1. Setup Environment

cd python-debug-ai-agent
python -m venv venv
# Windows:
venv\\Scripts\\activate
# Linux/Mac:
source venv/bin/activate
pip install -r requirements.txt

2. Dataset Generation

# Add API keys to .env
python -m data_pipeline.data_scraper
python -m data_pipeline.github_issue_scraper
python data_pipeline/data_cleaner.py
python data_pipeline/dataset_formatter.py

3. Train SLM

accelerate launch training/train_slm.py

4. Run Agent API

uvicorn app.api_server:app --reload --host 0.0.0.0 --port 8000

5. Test API

curl -X POST "http://localhost:8000/debug" \
  -H "Content-Type: application/json" \
  -d '{
    "code": "print(undefined_var)",
    "traceback": "NameError: name 'undefined_var' is not defined"
  }'

Dataset Strategy

• Sources: StackOverflow API, GitHub Issues API
• Size: 3000+ samples (train/val/test split 80/10/10)
• Format: JSONL with prompt-response pairs

Model Selection

• Primary: TinyLlama-1.1B (SLM, efficient)
• Comparison: Mistral-7B (larger baseline)

Fine-tuning Approach

• Technique: QLoRA (4-bit quantization)
• Hyperparams: lr=2e-4, epochs=3, batch=8, max_seq=512

Agent Architecture

User Input → Preprocessing → SLM Inference → Tool Chain → Structured Output
Tools: AST Parser, Static Analyzer

Limitations

• Scraping requires API keys (StackOverflow/GitHub)
• Training needs GPU (A100 recommended)
• Fix suggestions are suggestions (human review needed)