Syed Sabbih Haider Shah

Led validation of Hyper-CONUS v2 against HRRR across 2024 benchmarks, demonstrating lower RMSE for temperature and dew point. Directed design and delivery of a reproducible verification framework across 702 matched initialization times.

Train and evaluate spatiotemporal weather models across sub-seasonal, seasonal, and decadal horizons, including billion-parameter research workloads on Perlmutter and Frontier. Built peril-specific U-Net workflows and a 50,000-year thunderstorm-wind catalog.

Managed a team of 5+ and built the production AI inference stack for multimodal storytelling, including language, vision, and controllable diffusion models. Automated repetitive steps in short-form episode creation, reducing production time by about 40%.

Trained and deployed a custom Italian speech-recognition model achieving 4.1% word error rate and a punctuation model achieving 92% accuracy. Optimized inference for low-compute environments and implemented scalable audio pipelines.

Hired and managed a 15+ person team of researchers, MLOps engineers, mobile developers, and software developers; owned cross-functional ML projects from requirements through deployment. Built and deployed a 30,000-class LEGO recognition system with 96% top-3 accuracy.

Led an RGB-only facial anti-spoofing system from research through deployment, achieving 97.6% accuracy without depth input. Built and optimized real-time NVIDIA DeepStream pipelines, custom plugins, and CUDA-based GPU components.

Reduced model inference and deployment spend by 70% while implementing automated deployment, data logging, and metric monitoring. Deployed revenue forecasting at 4% mean absolute error and course classification at 96% precision and 95% recall.

Pursued a Master of Engineering in Machine Learning and Image Processing at Air University from 2017 to 2019.

Earned a Bachelor of Electrical Engineering (Electronics) from Air University from 2013 to 2017.