Key Notes
Production-ready model releases move through training, evaluation, optimisation and deployment readiness in this applied AI Performance role. Essential experience covers constrained production systems, hands-on PyTorch training and at least one stack such as TensorRT, CUDA, Qualcomm QNN, Triton or OpenCL. The work connects trained model behaviour with practical runtime constraints and on-vehicle deployment readiness.
What You'll Work On
- Own model releases from initial requirements through training, evaluation, iteration and final readiness for on-vehicle deployment.
- Train and refine deep-learning models in PyTorch through hypothesis-led experiments, ablations and explicit evaluation criteria.
- Diagnose production model-performance regressions, identify their root causes and apply quantisation, distillation or other optimisation methods when the trade-offs fit.
- Coordinate deployment hand-offs with adjacent ML and performance-engineering teams around bottlenecks, timelines, optimisation priorities and readiness criteria.
Why This Role Matters
- Wayve’s on-vehicle models must meet product and runtime constraints; the role’s training, evaluation and optimisation work determines which models are ready for deployment.
- OEM release cadence depends on models moving from training through evaluation and hand-off with explicit readiness criteria and identified performance regressions.
What They Are Looking For
- Experience: Improving production-system performance under latency, memory, bandwidth, power, thermal or cost constraints is essential.
- Technical: Hands-on training and iteration of deep-learning models in PyTorch is essential.
- Technical: Strong proficiency with TensorRT, CUDA, Qualcomm QNN, Triton, OpenCL or a comparable toolchain is essential.
- Technical: Reasoning from model behaviour to kernel, runtime and latency implications is essential.
- Preferred: Edge, embedded or real-time model delivery and device-level benchmarking are desirable.
- Technical: Familiarity with quantisation or distillation concepts is essential; hands-on application is a strong signal but not a strict requirement when fundamentals are solid.