About

At TCS Research in Kolkata, my main project is a bridge between the general neuromorphic frameworks and the hardware-specific ones. Take a spiking model written in SNNTorch and get the same behaviour out of Rockpool, Lava or Sinabs.

That sounds like plumbing, and a lot of it is. But every framework holds its own opinion about what a timestep means, when a membrane resets, and how much precision a weight deserves. Those opinions are buried in the model definition instead of written down. So the job became a more interesting one: state the network once, lower it onto each backend, then measure exactly where the spike trains stop agreeing. Ported time-series classifiers currently sit at 96% accuracy, within 2% of the hand-written versions.

Getting there meant learning SNNs properly rather than as an analogy. How event-based data should be represented before you hand it to anything. Which multimodal tasks genuinely get cheaper when computation is sparse, and which ones only look like they should.

Two things have my attention at the moment. The first is language. Text is the least natural fit for spiking networks: there is no event stream to begin with, the dependencies are long, and the embeddings are dense. I am trying to locate the real failure points rather than push a number up. The second is adaptive-LIF neurons, whose firing threshold moves as they spike. That gives a network a kind of short-term memory the weights never see, and I want to know how much of it can be trusted, and whether it belongs in the porting project.

Before Kolkata

M.Tech at IIIT Delhi, 2023 to 2025, AI specialisation, alongside two years as a teaching assistant for NLP, DSA, AI and Intelligent Systems. Nine months of that was at MIDAS Lab with Dr. Rajiv Ratn Shah, working on crowd analytics, person attribute recognition, and accident detection from live video, the last one jointly with TU Darmstadt. A final semester project on controllable summarisation with Dr. Shad Akhtar pulled me towards language models.

Before that, four years of Information & Communication Technology at Adani Institute in Ahmedabad. Signals and circuits, mostly, at a point when I had no idea any of it would be useful.

What I actually enjoy

The part where a model has to leave the notebook. A pipeline that must keep up with a video stream. An audio dataset that has to be recorded and labelled before a single experiment can run. A network that has to be re-expressed in a framework that never imagined it. That is usually where the real result is hiding, and it is almost never the part that gets written up.