At Turing, I produced rubric-based preference data for Google’s RLHF post-training pipeline and analyzed failure modes to inform data fixes. At CNTXT AI, I worked on multimodal training-data quality and evaluated medical and legal LLM outputs for hallucinations.
I build retrieval-augmented applications and guarded agentic workflows in Python. My projects include a multilingual SQL agent with schema-aware validation and a BookMind RAG pipeline with an evaluation harness.

