• Built a JIT-RAG pipeline that retrieves context (deploy correlation, anomaly cooccurrence, error spikes, semantic similarity search via pgvector embeddings) before invoking the LLM - reducing MTTR from hours to under 5 minutes (~95% reduction). • Implemented vector embeddings with pgvector cosine similarity search over incidents and commit diffs - powering a trace-to-commit correlation engine that semantically ranks the most relevant code changes behind an incident. • Orchestrated multi-step LLM chains with structured system prompts and enforced JSON schemas, including hallucination-prevention rules requiring evidence-backed citations for every diagnosis.