I am a postdoctoral researcher at the Ubiquitous Knowledge Processing (UKP) Lab, TU Darmstadt, joining in September 2026.

My doctoral research was on neural topic models and representation learning over document corpora. Since then I have worked on retrieval-augmented generation and on mechanistic interpretability for language models — circuit discovery through edge attribution and activation patching, and sparse autoencoders for model steering.

Most recently, in industry, I built an LLM pipeline that extracts compliance obligations from legislation: bare acts together with their densely cross-referencing rules and regulations. The hard part there was multi-document reasoning with provenance — following reference chains across documents and keeping every output traceable to its source under lawyer review. That question, how to make a model's reasoning something you can actually audit, is what carries over into my postdoctoral work.

I hold a Ph.D. in Computer Science from the School of Mathematical & Computational Sciences at IACS, Kolkata (2020–2025). I defended my thesis, Studies on Neural Topic Models for Document Corpora, on 15 January 2026, supervised by Dr. Debarshi Kumar Sanyal.

RAGAgentic AIMechanistic interpretabilityLLM safety & alignmentTopic modelingNLP

For more detail see my CV, or write to me.