wayve - Machine Learning Engineer, Synthetic Data

• Model work: Post-train or ablate a GAIA rig-transfer or pose-transfer checkpoint (NVS warp, calibration/intrinsics, shortcut/distillation). Inspect failures: black margins, odometry bias, flickering, wrong curvature column. • Generation at scale: Kick off SDS / Flyte jobs for tens of thousands of segments; debug GPU capacity, KV cache, DDIM step count, MCAP/delta-table correctness. • Landing in training: Binarise synthetic into a corpus-compatible table, wire sampling hooks, train RL and read suite + on-road diffs. • Capability expansion: Pose-transfer ELK/AEB sets, dashcam→Alpha3, or a new OEM rig (Nissan Proto 2.1, Gen3, BMW). • Cross-team: Driving-model owners on mix ratios, closed-loop/eval on whether generated MCAPs are usable as suites. • Paper / research alignment: Video generation, novel-view / camera transfer, flow-matching / distillation, data-centric training — only insofar as it changes a checkpoint or a mix we ship. • 🧩 What you’ll be working on • Post-train and iterate GAIA-class world models for synthetic-data capabilities: rig transfer (new camera/vehicle embodiments), pose transfer (rewritten ego trajectories), and related conditioning (geometry, calibration, actions). • Own the generation loop: config → large-scale GPU inference → training-ready artefacts, with clear lineage from the model and settings that produced them. • Land synthetic data in driving-model training (behaviour cloning, reward models, RL): binarisation, mix ratios, quality filters, and experiments that measure suite and on-road impact — including when synthetic should replace scarce real rig data. • Diagnose and fix geometry, calibration, and controllability failures (intrinsics/extrinsics, NVS warps, odometry/curvature, flickering, camera-layout artefacts) that determine whether generated video is training-grade. • Improve throughput and yield: inference optimisations (shortcut, distillation, KV cache, step count), valid-generation rate, and self-serve workflows so model developers can request synthetic sets without a specialist. • Expand coverage to new vehicle platforms and safety-critical scenarios (OEM bring-up; Emergency Lane Keeping / Automatic Emergency Braking). • Partner with world-model researchers, infra, and driving-model owners so generation, evaluation, and training stay one system. • 🙌 You should apply if • 4+ years in applied ML / research engineering, with a track record of training and shipping neural nets, not only operating data platforms. • Strong Python and PyTorch (or equivalent); comfort with GPU training, debugging, and reading model code. • Hands-on experience with video, generative, or world models (diffusion / flow-matching / autoregressive video, novel-view synthesis, neural rendering, or similar). • Working knowledge of cameras and 3D geometry (multi-camera rigs, intrinsics/extrinsics, warps/reprojection) and why they break generation or downstream training. • Evidence of taking generated or simulated data into a trained downstream model and measuring impact (mix, ablations, failure analysis). • Ability to operate generation or training at real scale (multi-GPU jobs, workflow orchestration, large video artefacts) and to make that path reliable. • Collaborative, experimental working style with researchers and platform engineers; you will own a capability, not a ticket queue. • 🌱 Not ticking every box? That’s totally okay! If you’re passionate about autonomy and keen to learn, we encourage you to apply even if you don’t meet every requirement.