Senior / Staff Research Engineer, Generative AI Deployment

About the role

Waabi, a Physical AI company behind commercial autonomous trucks and robotaxis, is hiring a Research Engineer focused on Generative AI Deployment in Toronto. You will own the path from research to production for the neural and generative models that power Waabi World, a multi-sensor simulator, working with teams on feed-forward scene reconstruction, diffusion and vision foundation models.

Key responsibilities

• Turn research code and checkpoints into versioned, reproducible, exportable production models that run in closed-loop simulation alongside GPU renderers and the autonomy stack.
• Profile models end to end and apply techniques such as model compilation (TensorRT), mixed precision, quantization, batching, CUDA graphs and custom kernels to meet latency and memory budgets; for diffusion and flow-matching models, also use faster sampling and distillation.
• Build benchmarks and numerical-parity tests for simulation realism and downstream autonomy metrics, and integrate regression checks in CI.
• Design export, versioning, model registry and build tooling so the simulation team ships new models reliably.
• Work with research scientists early so new architectures are exportable, efficient and hardware-friendly.

Required profile

• Experience shipping production-quality software at scale, including ML systems relied on daily by engineers and researchers.
• Hands-on experience taking models from research code to optimized production artifacts, measuring speedup and accuracy change.
• Ability to read research code for modern architectures such as vision transformers, diffusion and flow-matching models, and understand compute and memory usage.
• Experience launching and debugging GPU jobs in the cloud (AWS, GCP or equivalent), including orchestration, containers and keeping CUDA, PyTorch and TensorRT versions compatible across a fleet.

Required skills

• TensorRT, torch.compile, torch.export, ONNX, post-training quantization, quantization-aware training and pruning
• Nsight Systems, Nsight Compute and the PyTorch profiler
• CUDA and Triton kernel writing or modification