research

research under prof. adithya vadapalli at IIT kanpur, working through the MPC stack — garbled circuits, oblivious transfer, and the protocols that make private computation actually work in practice rather than just in theory.

the core question we're working on is about efficiency: standard garbled circuit constructions are correct but the constant factors are large enough that real deployments are limited to fairly simple functions. there's a lot of recent work on reducing the overhead through better gate representation and optimized OT protocols.

the web app linked below (CryptoVault-MPC) is the live implementation of Yao's protocol and a few variants — you can run the millionaire's problem and other boolean circuits in the browser and see the garbled evaluation executing.

if you're doing research in this area and want to talk, reach out.

co-authored with pavitra kushwaha, running on IIT kanpur's HPC cluster. the project is a systematic benchmark comparing discrete diffusion language models against autoregressive baselines for hindi NLP — four tasks, eight models, five-step evaluation pipeline.

the models: SEDD, LLaDA, MDLM, D3PM (diffusion side) vs. four autoregressive baselines. tasks: text classification, NER, question answering, generation. the benchmark is being designed to control for tokenization and data formatting differences that make a lot of existing cross-architecture comparisons hard to interpret.

hindi is an interesting test case for this because the architectural tradeoffs that show up in diffusion vs. AR comparisons for english don't necessarily generalize — the morphology is richer, the word order is freer, and the tokenization decisions interact with the model architecture in ways that matter.

not published yet. faculty recommendation letters are the validation path for now.