At Binocs, I design and benchmark multi-agent RAG pipelines and agentic workflows for enterprise documents, including text, PDFs, and images. My page-wise attribution system achieved 97% citation accuracy.
I also developed a Reflection Agent that self-scores responses against ground-truth criteria, and query decomposition and intent-routing modules for complex requests. I monitor agent quality with LangSmith and built summarization pipelines for document sets exceeding 100 pages.
During my AI Engineer Intern role at Sony Research India, I improved a customer churn model's F1-score from 0.620 to 0.637 and fine-tuned a language model on domain-specific storyline data. I also designed a rating-informed story generation system grounded in character details.
My research includes a generalized Direct Preference Optimization algorithm for LLM alignment, which demonstrated superior alignment across 3+ human preference benchmarks. I hold an M.Tech in Artificial Intelligence from Indian Institute of Science, Bangalore.
