軟體工程師-AI / ML Engineer (ADAS Full-Stack Machine Learning)-F76

【About the Role】

We are building an AI-centric ADAS platform supporting L2++ to L4 autonomous driving through a dual-path architecture: a classical ADAS pipeline for deterministic safety-certified behaviors, and a VLA (Vision-Language-Action) end-to-end model for open-world long-tail scenarios.
As an AI/ML Engineer, you own the full ML lifecycle — perception models, driving foundation models, and the data factory that feeds them — working primarily with raw sensor data from fleet logs. You ship models that run on the vehicle and build the tooling that keeps them improving at scale.

【Your Responsibilities】

Perception Modeling

•Design, train, and validate perception models: object detection, segmentation, lane and free-space estimation, occupancy prediction, and multi-camera depth estimation
•Develop BEV models that fuse multi-camera streams into spatial representations for downstream planning
•Define model I/O contracts, preprocessing and postprocessing boundaries with other engineers

VLA & End-to-End Driving Models

•Research and iterate on end-to-end driving models that produce trajectory proposals, waypoints, hazard scores, or action tokens from raw camera inputs
•Develop and evaluate VLA architectures: encoder-decoder designs, language-
conditioned policies, and imitation learning on large-scale driving data
•Build closed-loop evaluation pipelines: simulation-based testing, distribution shift analysis, and failure mode characterization
•Design arbitration interfaces between classical pipeline and VLA outputs
•Track VLA, world model, and autonomous driving foundation model research and assess production applicability

Data Factory AI

•Build auto-labeling pipelines using perception models and ensemble strategies to generate labels at scale across heterogeneous datasets
•Develop data mining models that surface hard cases, long-tail scenarios, and safety-critical events from fleet logs
•Build data quality scoring systems to gate label accuracy before training
•Design active learning workflows connecting model failure detection to targeted data collection
•Develop or integrate synthetic data pipelines for rare scenario coverage

Cross-Cutting

•Maintain model training infrastructure: distributed multi-GPU training, experiment tracking, and hyperparameter management
•Own ML evaluation frameworks: per-scenario metrics, regression tracking, and safety-critical recall monitoring
•Optimize all model types for embedded deployment via quantization, pruning, ONNX, and TensorRT
•Drive ML safety practices: OOD detection, uncertainty quantification, and robustness under sensor degradation
•Evaluate third-party AI solutions for perception, planning, and data tooling — assess technical fit against our architecture, define integration requirements, and own the technical collaboration with external vendors through to ECU integration

【You Have】

•Master's or above in Computer Science, AI, Electrical Engineering, or related field
•At least 3 years in production ML with exposure to at least one of: perception modeling, end-to-end driving models, or data pipeline ML
•Hands-on experience with large-scale distributed model training on GPU clusters
•Strong PyTorch proficiency; familiarity with ONNX export
•Deep understanding of modern deep learning architectures such as Transformers and Diffusion
•Solid experience with model compression and deployment: TensorRT, cuDNN, quantization-aware training, distillation
•Proficiency in Python across the full ML engineering stack
•Basic C++ proficiency for reading and debugging inference code
•Experience with dataset engineering at scale: annotation pipelines, quality control, and long-tail sampling
•Ability to communicate complex ML technical requirements clearly with external vendors and cross-functional partners in English