Built a mental-health intake copilot using RAG (LangChain + FAISS), automated LLM ops pipelines, & safety/prompt-injection auditing to ensure grounded, compliant responses. Fine-tuned LLaMA with SFT for tone & DPO for alignment & applied activation engineering (SAEs/steering) for fairness & hallucination reduction at inference. Designed schema-constrained AI agents to support routing decisions, improving accuracy & reducing manual review workload