About the job
NVIDIA's AV Eval team is building the next generation of driving behavior evaluation — moving beyond hand-crafted rules to learned evaluation using LLMs, VLMs, and agentic workflows. You'll define how we measure whether an autonomous vehicle drives well, building systems that bridge ML research and production evaluation. You'll ship systems that run at scale on real-world driving data and produce metrics that block or green-light software releases. In this role you will get to work on next-gen AV evaluation and create a direct impact on vehicle safety and shipping decisions. Join a new team being built from scratch — high ownership, high visibility to NVIDIA AV leadership
Responsibilities
Design and build learned evaluation pipelines that assess driving behavior using LLMs, VLMs, and multimodal models
Develop agentic workflows that chain model inference, retrieval, and structured reasoning to evaluate complex driving scenarios
Define evaluation-of-evaluation methodology — how do we know our learned evaluators are correct?
Build golden-set frameworks and calibration loops for learned metrics
Partner with AML (Alpamayo Logos) teams on model-specific eval needs (e.g., COT prediction quality, AML regression coverage)
Instrument evaluation systems with robust experiment tracking, A/B comparison tooling, and model versioning
Contribute to the team's transition from rule-based to learned evaluation: identify metrics and analyzers that are candidates for ML replacement and build the alternatives
Qualifications
Minimum
PhD with 4+ years, MS with 6+ years, or BS (or equivalent experience) with 8+ years of relevant experience in Computer Science, Computer Engineering, or a related technical field.
Hands-on experience building LLM/VLM-based pipelines — fine-tuning, prompt engineering, retrieval-augmented generation, chain-of-thought
Track record of shipping ML systems to production (not just prototyping or publishing)
Strong software engineering fundamentals — you write clean, tested, reviewable code in Python and C++
Experience with evaluation methodology: precision/recall, inter-rater reliability, calibration, annotation pipelines
Comfort with large-scale data processing (Spark, Dask, or similar)
Strong Python skills. Experience with PyTorch or JAX. Comfortable with GPU-based training workflows.
Preferred
Autonomous driving, robotics, or safety-critical domain experience
Familiarity with driving behavior taxonomies (cut-ins, hard braking events, lane-keeping metrics, scenario-based evaluation)
Experience with video understanding models or multi-modal evaluation. Knowledge of agentic AI frameworks (LangChain, DSPy, CrewAI, or custom)
Track record of influencing technical direction across team boundaries
Experience with LLM/VLM fine-tuning or application development