BMW Group × MIT GenAI Lab
Continuous Evaluation & Autonomous Prompt Optimization
We are building a closed-loop system that improves AI agents without retraining models. Instead of fine-tuning, we treat the system prompt as a policy layer and optimize it using structured evaluation feedback.
Loop:
Evaluate → Modify Prompt → Re-test → Select Improvement
Goal: enable agents to adapt to evolving tasks in non-stationary production environments.
- Automated prompt optimization loop
- Measurable improvement over baseline
- No model weight updates
- No RL or fine-tuning
For now, proof-of-concept only.
Google Colab
Used for compute capability. We use our student accounts.
GitHub (this repo)
Used for:
- Centralized version control
- Benchmark tracking
- Team visibility
Colab = compute + coding
GitHub = storage + record
Data/
Evaluation datasets.
Experimentation/
Individual research notebooks.
Notebooks/
Versioned system pipeline iterations. This is what we track and share; the final polished notebooks.
Outcome: a self-evolving prompt optimization pipeline for enterprise agents.