- Developed in-house ML model for multi-lingual low-latency speech-to-translated-text- Improved ML model inference efficiency by 5x to support scaling demand and reduce costs by 80%- Implemented modular streaming framework to provide continuous stateful inference with auto language-detection and code-switching support, backed by stateless ML models that can easily be hot-swapped and scaled through CI/CD- Developed API service to support live streaming speech-to-translated-speech with zero-shot voice cloning over WebSocket/Socket.IO connections

- Managed overseas engineering and data science teams- Designed no-code self-service AI software suite to allow users to explore, analyze, and experiment with their own data

- Created legacy document digitization framework for automated parsing and de-duplication of semi-structured data in PDF, Word, and Excel files, reducing manual work from months to minutes- Developed distributed OCR platform to digitize millions of savings bonds autonomously in hours

- Mentored undergraduate research intern during 2016 SULI program (work presented at APS-DPP 2016)- Designed and conducted experiment to validate tethered spring-mass model of magnetorotational instability in a lab setting for the first time- Created an automated motion tracking script for analyzing 200GBs of GoPro camera data- Analyzed test mass trajectory, and verified results with simulations of theoretical model in Python

- Visualized performance of dielectric waveguides and coupled-cavity traveling-wave tube (TWT) amplifiers in parameter space to investigate unwanted wave oscillations- Analyzed dispersion relations of TWTs to determine threshold parameters for onset of absolute instability; Results published in Physical Review Letters- Constructed probabilistic models to study I-V asymmetry in 2-D funnel-shaped ballistic diodes, and verified analysis with Monte Carlo simulations using MATLAB; Results published in Journal of Applied Physics- Derived generalized 3-terminal thin film contact resistance scaling law from regression models in Python with MAPE <0.5%, valid for parameters spanning 4 orders of magnitude, and validated results against numerical simulations in ANSYS Maxwell; Results published in Journal of Physics D: Applied Physics

- Analyzed effects of existing infrastructure on final proton beam spin polarization at Fermilab- Optimized parameters of superconducting solenoid addition to improve 8.9 GeV/c beam line final spin polarization from 83% to 98%- Developed beam parameter optimization code to improve 120 GeV/c beam line final spin polarization from 65% to 98%, eliminating need for $2M spin rotator addition in SPIN@FERMI collaboration