At VideoDB I built a real-time fact-checking system for live audio and video streams. Given a stream from YouTube Live or Google Meet, the pipeline transcribes claims as they are spoken, retrieves indexed video evidence, and decides within a second whether each claim is grounded in what the camera actually shows.

I also open-sourced VideoDB Claude Skills, a small library exposing video search, transcription, and scene understanding through a plain-English terminal interface using the MCP protocol. It shipped to the Claude developer marketplace.

I designed and optimised large-scale multilingual ASR and speaker diarization pipelines over GigaSpeech 2, Emilia, and NVIDIA Granary, analysing trade-offs in normalization strategy, language mixing, and benchmark generalization across regional accents.

My current paper builds on Indian-ASR-Bench, a WER benchmark of five ASR systems on the TIE dataset — 986 Indian English academic lecture clips. I evaluated Whisper Base, Medium, and Large alongside Parakeet-TDT-0.6B and Qwen3-ASR-1.7B, with breakdowns by region, speech rate, audio duration, gender, and discipline.

The first project was PustakAI, a curriculum-aligned QA pipeline for Indian school textbooks. I built the curation and evaluation harness, designed RAG-based prompting strategies, and fine-tuned Gemma3 :1B and LLaMA3.2 :3B variants. The paper was accepted at ACM COMPUTE 2025. The NCERT dataset is public on HuggingFace.

The second project is ICH-QA, a 131,495-pair synthetic QA dataset over 7,506 Wikipedia articles on Indian Cultural Heritage. RAG reaches 48% EM - confirming the dataset requires genuine cultural grounding, not surface pattern-matching.