Senior AI/ML Engineer

GM•Sunnyvale, CA

6d•Hybrid

About The Position

The Compression and Parity team in GM's Autonomous Vehicle organization is responsible for ensuring model optimization is safe enough for repeated deployment. This team compresses and quantizes models for use in vehicles and maintains the analytical tools that verify the compressed models' behavior against the original, providing auditable evidence. The tooling developed by this team is critical for every deployment decision regarding compressed models within the organization. This role focuses on model numerics, requiring an engineer with strong mathematical rigor to measure, bound, and reason about the numerical behavior of shipped models, and to translate this analysis into deployment decisions. The core challenge is determining if two numerically different versions of the same model are safe, which involves connecting disparate fields like floating-point drift and Hessian conditioning with vehicle trajectory error. The engineer will develop the quantitative tooling to establish this connection and define the thresholds for ship/no-ship decisions. This is not a role for simply running existing validation harnesses; the expectation is to identify the specific operations causing divergence and the underlying reasons, leveraging existing tooling to expedite this investigation.

Requirements

Working command of numerical analysis and matrix theory (Jacobian/Hessian estimation, spectral properties, conditioning, floating-point error analysis).
A strong mental model of neural network training (loss landscapes, gradient/error propagation, optimizer dynamics, training failure modes).
An adversarial instinct for numerical edge cases (denormals, extreme dynamic range, catastrophic cancellation, degenerate shapes, accumulation-order effects).
High proficiency in PyTorch and Python.
Track record of building analytical tools that are mathematically defensible and production-ready.
Judgment to make analysis actionable (identifying key quantities, tracing anomalies to operations).
Bachelor's, Master's, or PhD in Applied Mathematics, Control, Physics, Computer Science, Data Science, or a closely related quantitative field.

Nice To Haves

Hands-on experience debugging large-scale training runs: diagnosing loss spikes, resolving numerical divergence, and investigating training dynamics.
Experience with distributed training beyond DDP (FSDP, Megatron-LM, DeepSpeed, 3D parallelism), especially designing low-overhead observability over sharded parameter and gradient state.
Experience in quantization, compiler toolchains, or inference-time numerical parity.
Experience building or scaling evaluation pipelines and metrics formulation for AV or ADAS systems.
Published or applied work in model robustness, OOD generalization, or adversarial/perturbation analysis.

Responsibilities

Validate Optimized implementations: Establish rigorous equivalence between optimized and reference implementations, defining what equivalence means and designing tests to expose violations.
Connect tensor differences to behavioral disparity: Map low-level numerical differences (quantization, compilation, precision reduction) to downstream driving behavior using open-loop metrics (e.g., trajectory displacement error, perception IoU) and closed-loop outcomes, identifying the causal mechanisms.
Build sensitivity and robustness analysis tooling: Characterize model output response to weight and input perturbation using Jacobian/Hessian-based methods, extend this to out-of-distribution inputs, and create reusable tooling for engineers.
Build training dynamics observability: Design diagnostics to detect and root-cause training instabilities (gradient vanishing/explosion, loss spikes, silent divergence), including decomposition of gradient and update trajectories.
Make it cheap enough to always be on: Ensure metric computation, gradient decomposition, and diagnostic logging run within real distributed training jobs with negligible throughput cost and no OOM risk.