San Jose, California, United States
I build interpretable evaluation for transformer-based video/image/text/tabular models — metrics, MLOps pipelines, and adversarial failure harnesses. PhD Concordia 2025; 12 papers (ICSE, IEEE TCC, ACM TOMM). What I do: • Scalable automated eval pipelines for visual / multimodal generative models — microservice architecture, OpenAPI contracts, deployed on Azure / GCP / AWS with regression testing, dashboards, and drift monitoring. (XAIport, ICSE 2024 A*) • Perceptual quality metrics for video transformers — faithfulness, monotonicity, temporal consistency. • Failure-case discovery + regression-defense harness, first joint spatio-temporal adversarial attack on video transformers. • 3D vision and evaluation; Gaussian Splatting reconstruction from phone video; metrics for geometric consistency. • Production multimodal AI in flight — vision encoders + LLM orchestration, cross-modal validation, hallucination-detection layers.
Shipping production multimodal AI pipelines for architectural design generation. • Multimodal generative pipelines — vision encoders + LLM orchestration • 3D vision in production — SAM3D for 3D segmentation; Gaussian Splatting / NeRF reconstruction from phone video; geometric and multi-view consistency evaluation for generated 3D assets. • Camera rendering pipelines — chaining vision encoders with generative decoders for real-time architectural scene synthesis under production SLA. • Real-time AI chat with context-aware multimodal responses driving iterative design loops. • Stack: Python, PyTorch, microservices, OpenAPI, cloud deployment.
• Built and ran numerical simulations of fluid transport / multi-physics coupling / dynamical systems / agent-based models] using Python / MATLAB / COMSOL / OpenFOAM. • Implemented finite-volume / finite-element / Runge-Kutta / spectral solvers for Navier-Stokes / advection-diffusion / control-system equations; validated against analytical benchmarks and experimental data. • Early exploration of data-driven surrogates for expensive simulations — set up the conceptual bridge tolater PhD work on transformer-based modeling of spatio-temporal phenomena.
Simulation of a steam system within UniSim Design (sequential algorithm) and a new Unisim equation oriented solution environment including error description and reporting Simulation of a steam system within Linde’s inhouse simulation environment (Optisim) and comparison of results Use the simulate results to solve design problems for customers Participation in P&ID review meetings