Add ART email search agent project

[strickvl]

Summary

This PR adds a new project demonstrating how to train an email search agent using OpenPipe ART (Agentic Reinforcement Training) with ZenML for production ML pipelines.

• Train an RL agent using GRPO (Group Relative Policy Optimization) with RULER scoring
• Track artifacts including scenarios, model checkpoints, and training metrics
• Orchestrate on Kubernetes with GPU step operators for training
• Evaluate models with automated correctness judging

The agent learns to search through emails and answer questions using LangGraph's ReAct pattern, starting from a Qwen 2.5 7B base model.

Architecture

data_preparation_pipeline (cached, no GPU)
  download_enron_data → create_database → load_scenarios

training_pipeline (GPU required)
  setup_art_model → train_agent (LangGraph rollouts + RULER + GRPO)

evaluation_pipeline (GPU required)  
  load_trained_model → run_inference → compute_metrics

Test plan

• Verify data preparation pipeline runs locally
• Test training pipeline with small config on GPU
• Validate evaluation metrics computation
• Test Kubernetes config on cloud cluster

[dagshub]

Join the discussion on DagsHub!