Reducing hallucinations in LLMs through mechanistic interpretability
Addressing hallucinations in large language models through mechanistic interpretability — dissecting neural network internals to uncover why models produce factually incorrect outputs. Awarded grants from VCs and ExceptionRaised.
Semantic search & retrieval for IIT Madras historical archives
A scalable system that indexes and retrieves historical records from the IIT Madras Archive Center using embeddings, OCR, knowledge graphs, and frameworks like LangChain, GraphRAG, and FAISS.
Speaker diarization with emotion & language detection for Indic languages
An advanced speaker diarization system with integrated emotion and language detection, designed to handle complex code-mixed speech and support multiple Indic languages — one of the most challenging problems in neural audio processing.
