Staff Machine Learning Engineer - Vision-Language Foundation Models

The Team & Mission: In the Oracle Perception team, our mission is to build the ultimate cognitive engine for autonomous driving. We are pioneering the use of large multimodal foundation models (e.g., Gemini) to build a powerful offboard reasoning and data flywheel system. We are moving beyond traditional perception to true scene understanding and driving actions—building offboard models that can comprehend complex driving problems, predict object/scene dynamics, and deduce driving paths with logical rationale. Our core focus is advancing the VLM foundation itself. By pushing the boundaries of multimodal pre-training and state-of-the-art post-training (SFT, RL), we are creating models capable of rich, reasoning-based autolabeling at a massive scale. This closed-loop data engine directly powers the training and evolution of Waymo's real-time onboard models.

Responsibilities

Drive Pre-training & Domain Adaptation: Lead the technical strategy for curating and constructing massive-scale, high-quality multimodal pre-training datasets. Define data mixture strategies to instill deep, Waymo-specific driving intuition and physics-grounded understanding into foundation models without catastrophic forgetting; Lead Post-Training & Reasoning Enhancement: Design and implement state-of-the-art fine-tuning (SFT) and Reinforcement Learning (RLHF/RLAIF, DPO/GRPO/PPO) pipelines. Drastically improve the model’s instruction-following and complex reasoning capabilities (e.g., Chain-of-Thought, spatial-temporal reasoning, and driving rationale prediction); Pioneer the VLM Data Flywheel: Architect the highly scalable inference and evaluation pipelines that leverage these trained Gemini-class models to autonomously source, sample, and autolabel critical edge cases, directly accelerating the onboard perception models; Define Training Recipes & Scaling Laws: Conduct rigorous ablation studies to optimize model architectures, token budgets, and loss functions. Establish best practices for scaling multimodal training efficiently on large GPU/TPU clusters; Drive Cross-Functional AI Strategy: Act as the principal technical visionary across ML Infra, Perception, Behavior, and AI Foundation teams. Drive consensus on the data flywheel architecture and embed VLM reasoning capabilities seamlessly into the broader autonomous vehicle stack; Provide Staff-Level Technical Leadership: Own the long-term technical roadmap for foundation model development. Mentor senior engineers, lead rigorous design reviews, and establish standard-setting engineering practices from advanced prototyping to production deployment

Requirements

Master’s degree in Computer Science, AI, ML, or a related technical field; 8+ years of hands-on experience designing, training, and scaling deep learning models, with at least 3+ years focused deeply on training Large Language Models (LLMs) or Vision-Language Models (VLMs); Proven expertise in the full lifecycle of Foundation Models: from pre-training data curation (interleaved formats, tokenization) and distributed training to advanced post-training techniques; Expert-level understanding of training infrastructure and distributed paradigms (e.g., FSDP, Megatron, JAX/Pax) required for training massive models reliably; Expert-level software engineering fundamentals using Python, PyTorch, or JAX, with a track record of building reliable, highly scalable ML systems; Proven ability to operate with high ambiguity, define technical roadmaps, and drive complex, multi-quarter technical initiatives across multiple teams in a fast-paced environment

Nice to have

PhD in Computer Science, Artificial Intelligence, or a related field; Strong publication record in top-tier AI venues (e.g., NeurIPS, ICML, ICLR, CVPR) focusing on foundation models, large-scale training, reinforcement learning, or reasoning; Deep experience with advanced Reinforcement Learning paradigms applied to language or vision tasks (focusing on improving System 2 thinking, logical deduction, and model alignment); Demonstrated experience in Data Engineering for Foundation Models at the scale of billions/trillions of tokens (e.g., deduplication, quality filtering, synthetic data generation); Familiarity with the systemic challenges of multimodal perception in robotics or autonomous driving (e.g., 3D scene understanding, trajectory prediction); A proven track record of Staff-level impact: influencing product direction, pioneering zero-to-one ML architectures, and multiplying team efficiency through technical leadership

Benefits

Health and wellness; Financial wellness; Flexibility and time off; Top-notch medical, dental and vision insurance; Mental wellness support; Gym membership; Special wellness programs; Competitive compensation; Bonus opportunities; Equity; Employees provident fund; Employee discounts; Ability to work from another location for four weeks per year; On-site or hybrid work model; Remote working opportunities; Paid time off; Bereavement, sick, and parental leave