At Mercedes-Benz Research and Development India, I'm building scientific machine-learning surrogates for turbocharger compressor CFD. They predict full 3D flow fields and performance-map quantities at unseen operating points, with Gaussian-process uncertainty helping select new simulations.
I built a reduced-order modelling pipeline for transient compressor simulations, comparing POD and DMD with autoencoders. It cut CFD post-processing time by 50% while preserving dominant flow dynamics, and DMD extrapolation forecast future flow states at 2.5% relative error.
I designed and deployed a multi-agent diagnostic assistant for signal ranking, which cut root-cause analysis time by 40% across hundreds of investigations. I also integrated signal retrieval and historical RCA knowledge into a Databricks application.
For field-data analysis, I built PySpark pipelines processing up to 26 TB of vehicle sensor data and developed automated methods that cut RCA from about two days to three hours. I also led a team of three building a web framework for plausibility checks and anomaly detection, used by 20+ domain experts in beta.

