Responsibilities
- Own the perception model end-to-end: architecture, training, evaluation, and deployment. The core challenge is building a model that generalizes across geographies, road types, sensor setups, and environmental conditions without per-vertical forks.
- Drive a camera-first perception strategy. The goal is to progressively reduce dependencies on HD maps and lidar. How to get there is part of the job.
- Lead training and iteration cycles hands-on. You will be in the data, the eval dashboards, and the failure analysis. When perception regresses in a new geography or road type, you own understanding why and fixing it.
- Own model performance across the full deployment surface: highway, urban, residential, ramps, complex intersections, poor weather, hilly terrain. You care about on-vehicle driving outcomes, not just offline metrics.
- Manage the model lifecycle from training through quantization and deployment on embedded compute, including device-specific optimizations. Close the gap between what the model does offboard and what it does on the vehicle.
- Work directly with OEM customer programs to understand sensor configurations, target ODDs, and performance requirements. Translate these into model architecture and data strategy.
- Recruit, develop, and technically lead a team of perception engineers. Set high technical standards and create a culture of rigorous experimentation and measurement.
Requirements
- 5+ years in ML/deep learning for perception or 3D scene understanding. Deep hands-on experience training and deploying vision models at scale.
- 2+ years managing or technically leading a perception team, with ability to both set direction and contribute to architecture and training decisions directly.
- Experience building production perception systems, especially camera-only or camera-first solutions.
- Track record deploying perception models to embedded hardware under real-time latency and compute constraints, including device-specific optimizations.
- Strong software engineering in Python and C++, comfortable across the stack from training code to onboard inference integration.
- Experience scaling perception models across multiple geographies, sensor setups, or vehicle platforms.
Nice to Have
- Deep familiarity with transformer-based architectures for 3D perception, BEV representations, multi-task learning, and dense prediction.
- Familiarity with occupancy-based scene representations, sparse query-based architectures, or temporal aggregation approaches.
- Experience reducing or removing map dependencies in perception systems.
- Background in autolabel pipelines, data quality monitoring, or data flywheel design for perception.
- Experience with closed-loop simulation for perception model evaluation (neural sim, log sim, scenario-based testing).
- Experience at an AV company that has shipped perception to production.
