At Deccan AI, I evaluated Gemini-powered agentic systems for tool-call accuracy, trigger sequencing, and multi-step reliability. I also rated LLM outputs in RLHF pipelines and documented edge cases that informed model fine-tuning.
I built an ArXiv RAG question-answering system with explainable PDF ingestion, semantic search, and answer generation across Gemini and Groq. I added an LLM-as-judge layer to assess faithfulness and relevance, then used its scores to refine chunking and prompts.
For a customer support prioritization project, I combined email classification with rule-based urgency signals in a live Streamlit dashboard. I also designed a voice-guided form assistant, whose product spec won first position at a college-level hackathon for Swecha.

