# Track 15: Evaluation, Observability & Safety

The track that makes everything before it production-grade rather than a
demo. Evaluating a system whose output is free-form text is genuinely
hard and needs its own methods — eval datasets, LLM-as-judge, regression
testing prompts — plus hallucination detection, guardrails, agent
tracing, and keeping token spend from quietly exploding.

## Modules

1. Why GenAI Evaluation Is Hard
2. Building an Eval Dataset
3. Deterministic Evals and Assertions
4. LLM-as-Judge
5. Human Evaluation and Annotation
6. Benchmark Suites and Their Limits
7. Regression Testing Prompts
8. Hallucination: Causes
9. Hallucination: Detection Techniques
10. Grounding and Citation Verification
11. Guardrails: Input Filtering
12. Guardrails: Output Validation
13. PII and Data Leakage
14. Tracing Agent Execution
15. Metrics, Dashboards and Alerting
16. Cost Monitoring and Optimization
17. Red-Teaming and Adversarial Testing
18. Capstone Project