Job Description & Details
This is a hands-on contract role bridging the gap between theoretical AR/VR research and actual silicon. If you enjoy the pain and glory of taking bloated Python models and getting them to execute in real-time within strict power and memory budgets on embedded hardware, this is your playground.
What You'll Actually Be Doing
You'll spend your days side-by-side with research scientists who build incredible audio and computer vision models that unfortunately melt mobile processors. Your job is to translate their PyTorch prototypes into lean, mean C/C++ inference pipelines. You'll profile memory bottlenecks, apply aggressive quantization, and figure out why a specific operator is choking the DSP or NPU. Expect plenty of ambiguity, shifting requirements, and deep hardware-level debugging.
The Core Tech Stack
You need absolute fluency in C/C++ and deep learning frameworks like PyTorch or TensorFlow, but the real differentiator is your grasp of on-device inference constraints. Mastery of quantization techniques, pruning, distillation, and runtime frameworks like ExecuTorch or TFLite is mandatory. If you also understand audio DSP fundamentals or computer vision pipelines, you'll hit the ground running immediately.
Interview Expectations
Expect the hiring team to hand you a scenario where a model's latency is double the real-time threshold on a target embedded chip, and they'll want to hear your exact playbook for profiling and shrinking it. They are secretly testing whether you understand hardware-software co-design—meaning you can push back on researchers with architectural trade-offs rather than just blindly trying to optimize bad tensor shapes.
Application Advice
Your resume needs to scream embedded ML performance optimization, not just generic data science or web development. Explicitly highlight your production C/C++ experience alongside specific tools you've used for model compression and quantization. Make sure keywords like quantization, latency profiling, PyTorch, C/C++, and embedded inference are front and center so you easily clear the ATS.