At Learn2AI, I own an AI course assistant’s question-answering system, from data through deployment. I cut cross-domain retrieval errors from 18% to 2% and improved precision at rank one from 64% to 89% while keeping P95 latency below 200 ms.
At SJSU AI Labs, I designed TempRALM, a temporally aware retrieval-augmented language model that improved Recall@1 by 166% on more than 60,000 passages. The work received a Best Paper Award.
I also researched localization for low-resource languages, curating a conversational parallel corpus and fine-tuning mBART and LLaMA-2. My experiments included human-in-the-loop evaluation and ablations across data size, encoders, and model sizes.
Earlier, at Lotus Infosystems, I engineered Java and Spring Boot APIs and reduced P95 latency by 40%. At Sahyog, Centre for Promoting Health, I shipped an X-ray model pipeline with human review and created a dashboard with Grad-CAM explainability.
