Sepideh Shaterian

Ml Software Engineer @Cohere

Markham, ON, CA

Productionized NPU-accelerated inference by contributing to the ONNX Runtime Execution Provider used by Microsoft Copilot.• Contributed to Dynamic Dispatch (Vitis AI, open source): executes ONNX subgraphs as one NPU call for more consistent, faster inference.• Implemented dynamic-shape support so models handle variable input sizes without re-compilation, maintaining competitive throughput.• Added in-runtime CPU fallback and buffer reuse for mixed CPU/NPU paths, reducing peak memory.• Integrated a compiler flow (use prebuilt binaries when available; generate at runtime when not) to broaden operator/model coverage.• Built change-aware CI + a results dashboard to run only affected tests and visualize outcomes by model/metric, shortening feedback cycles and surfacing regressions early.