Performance Engineer, On-Device Inference

About the role

Take Sarvam models from research‑handoff state to production‑ready artifacts on at least two of our target chipsets (Intel xPU, ARM xPU, Apple xPU, Nvidia/AMD GPUs). You will own one to two (model, chipset) pairs end‑to‑end and work closely with the consuming application team during integration.

Key responsibilities

• Quantize models, validate accuracy, benchmark performance, and document results.
• Author a deployment workbook for each model‑chipset pair you own.
• Embed part‑time with consuming teams during integration, debugging performance and accuracy issues together.
• Maintain and extend the team’s benchmark harness.

Required profile

• 3+ years of experience building ML systems.
• Solid PyTorch and ONNX export experience, including handling dynamic shapes, control flow, and custom ops.
• Hands‑on quantization in production on at least one real model.
• Comfort with at least two of the following runtimes: ONNX Runtime, TensorRT, CoreML, OpenVINO, QNN, LiteRT.
• Fluency in profiling on at least one platform.

Required skills

• PyTorch
• ONNX
• Quantization
• ONNX Runtime
• TensorRT
• CoreML
• OpenVINO
• QNN
• LiteRT
• Profiling
• Custom op authoring