Hi, I'm Soumil. I'm a computer engineering student at Purdue. Right now most of my time goes to AI research at Perforated AI, where I work on dendritic model compression, and to the electronics on Purdue Formula SAE.
The rest goes to a habit I can't shake, which is taking a published result and checking it myself. Most of what's below came out of that.
Before Purdue I was in the TMS lab at Harvard Medical School, did research at IIT Roorkee, and wrote software at McKesson and Tekizma.
If you're building something nearby, I'd genuinely like to hear about it, so email me.
Selected work
- 2026deadboltSix published backdoor detectors, run blind against the same 102 models and scored on identical metrics. One attack beats all six: following any of them is worse than a coin flip.
- 2026nanoDLMThe smallest masked-diffusion language model you can actually read, in about 1,100 lines of plain PyTorch. Trains in 22 minutes on one 3090, against a matched autoregressive baseline.
- 2026BastionA local-first control plane between an agent and its MCP servers. It caps spend, blocks dangerous arguments and logs every call. Installable as bastion-mcp.
- 2026halluscopeThree published hallucination detectors, scored by AUROC on HaluEval and TruthfulQA, with a demo that paints the fabricated spans red. Runs on one GPU.
- 2026SecuriVaxA cold-chain monitor that tells you whether a vaccine is still good. Best Overall at HopHacks.
- 2026ATLASClosed-loop TMS threshold estimation, built in the Harvard lab. It scores motor evoked potentials with calibrated confidence and picks each next stimulus by minimising entropy.