Discrete diffusion for language
My work explores discrete text diffusion and Hindi NLP, including comparisons of SEDD and LLaDA-style models with autoregressive language models.
The questions start with representation: how Devanagari is tokenized, how corruption changes a sequence, and how denoising behavior shifts across languages. Comparing model families also means accounting for sampling steps, tokenization, and inference cost.
Read my Hindi diffusion noteMPC & cryptography
I’m a Research Fellow at IIT Kanpur, working on MPC and cryptography under Prof. Adithya Vadapalli.
Secure multi-party computation studies how parties can compute a joint function while controlling what is revealed about their inputs. The interesting boundaries sit in the protocol, the threat model, and the cost of communication.
Simulation infrastructure
At 
Vivacity, I work on the runtime connecting agents to executable environments. State, branching, and domain-specific verification are central to that work.
Papers behind the work
Background reading for the model families above.
Discrete Diffusion Modeling by Estimating the Ratios of the Data DistributionAaron Lou, Chenlin Meng, Stefano Ermon / ICML 2024Introduces score entropy for learning discrete diffusion models.
Large Language Diffusion ModelsShen Nie et al. / 2025LLaDA models language through a masking process and learned token prediction.