IIT Mandi - research intern

I spent May to September 2024 as a research intern at IIT Mandi, working on two projects. This was before Binocs, before ViewR, when I was still figuring out whether research was a fit for me. Two years on, most of the technical specifics are stale. What stuck is what I will talk about.

Project one was continuous authentication using behavioral biometrics - keystroke dynamics and touch patterns on mobile. The thesis was that traditional one-shot login (password, fingerprint) leaves a gap, and you could continuously verify the same person was still holding the device by their typing rhythm and swipe patterns. I built data collection tooling for an Android app and trained a small classifier in PyTorch. The model worked above chance, not at production accuracy.

Project two was a graph neural network for molecular olfaction - predicting what a molecule smells like from its molecular graph. This was based on Google's olfaction GNN paper that came out around the same time. I built the dataset pipeline (SMILES to molecular graphs using RDKit), implemented a message-passing GNN in PyTorch Geometric, and ran experiments on the public olfaction datasets.

What I will say honestly in interviews - the work was educational, not breakthrough. I learned PyTorch from being on the line for a real model. I learned that research moves at a different cadence than engineering (you can spend a week on a hyperparameter sweep that ships nothing). I learned that the bottleneck in research is often the dataset, not the model.

What stuck more than the technical content was the workflow. Read 3 papers a week, summarize them in the lab Slack, present one in lab meeting every other week. That discipline of reading and writing about other people's work is the part I still use today, even in product engineering.

Both projects were learning exercises. The continuous-auth model never shipped anywhere. The GNN reproduced known results but did not extend them in a new direction. I left in September to go solo at ViewR.