Research

My work asks what a model can reliably infer from sensors that do not measure the target directly. I apply it to wearable health and to industrial control, and two supporting lines of work, explaining sensor data in language and running models on small devices, make those models easier to check and to deploy.

Application

Wearable health sensing

Estimating physiological state that a wearable cannot measure directly, with models a clinician can inspect.

With the Kireev Lab at UMass Amherst, I estimate blood pressure from graphene electronic tattoo patches with physics-informed and tree-based models, to track blood pressure continuously without a cuff. Our PhysioNet Challenge 2026 entry screens for future cognitive impairment from overnight sleep studies using interpretable sleep features aligned to the scored hospital. On the hidden test set it scored 0.678 age-conditioned AUROC, 10th of 42 teams. Earlier, I predicted hypoxemia severity in hospital in-patients from the photoplethysmograph.

Industrial cyber-physical systems

Finding faults in industrial control software from the plant's sensor data, without instrumenting the controller.

With Bosch Research, I reconstruct the branch-level program execution trace of an industrial controller from the plant's process sensors alone, so faults in control software can be localized without adding logging to the deployed controller. On a Bosch ctrlX testbed it detects code-path faults that sensor readings alone cannot reveal at 0.99 AUC, and with a healthy replay as reference it names the faulty function 69% of the time. During my Bosch internship I built a contrastive model linking four modalities from an industrial control testbed (vision, sensor, program trace and text data) for language-guided few-shot state classification.

Supporting method

Explaining sensor data in language

Making a model's reading of sensor data inspectable by letting people ask about it in plain language.

LLaSA pairs an IMU encoder with a language model so it can answer open-ended questions about human activity from raw accelerometer and gyroscope data, and it outperforms GPT-3.5 and GPT-4o-mini on these tasks. We released the SensorCap and OpenSQA datasets, with 35,960 captions and 199,701 question and answer pairs. Concord studies when to trust LLM judges, using agreement among several judges as an uncertainty signal. This builds on two and a half years of production conversational AI and speech recognition for Bengali speakers.

Efficient inference on the device

Keeping these models within the battery and latency limits of wearables and embedded hardware.

JOLT trains early-exit networks so that each exit's confidence tracks which inputs it gets right, then routes on that confidence so that it overspends a stated compute budget by at most 6.8% when the input distribution shifts. Across seven sensor, image, audio and language datasets it ranks first of eleven methods. At WPI I was also a teaching assistant for the graduate course On-Device Deep Learning.