Rohit Mahesh

I study whether the same hardware or system fault can change a robot's action and fool the monitor meant to catch it. I'm building fault-injection and evaluation tools around OpenVLA, LIBERO, and SAFE, with a focus on reproducible experiments and systematic fault selection.

I've spent several years building machine-learning tools that turn microscopy heatmaps into quantitative measurements. My most recent project adapts a U-Net, runs it with ONNX Runtime, and gives nontechnical lab members a review interface; it cut each analysis from about 30 minutes to 3. We're now moving it onto MSU servers and preparing a manuscript.

Utero-ViT code WaveAtlas site WaveAtlas code

I developed a longitudinal medical-imaging pipeline for tuberculosis outcome prediction, extracting Vision Transformer representations from serial chest X-rays and modeling temporal disease progression with sequence-level aggregation.

code

I was a core contributor on a four-person team building Lenovo's cross-device AI assistant. I shipped features for a major Windows and Android release in Kotlin and Rust, worked across on-device inference and device state, and often became the first call when Product or Design found something broken.

I developed software for multiline tactile Braille displays, including an English-to-Braille translation engine, an accessible menu framework for tactile navigation, and low-latency audio cursor cues iterated with blind co-designers.