Senior AI/ML Engineer

Description

The Compression and Parity team sits inside GM's Autonomous Vehicle organization. It compresses and quantizes models bound for the car and builds the analysis that shows a compressed model still matches the original, using bounded, quantified and auditable evidence. Every deployment decision the organization makes about a compressed model goes through this team's tooling.

This role owns model numerics: measuring, bounding and reasoning about how the shipped models behave numerically, then converting that analysis into deployment decisions.

The core question is simple to state: given two numerically different versions of one model, is the difference safe? Answering it means linking topics usually studied apart, from floating-point drift and Hessian conditioning to vehicle trajectory error. You will build tooling that quantifies that link and set the thresholds behind ship or no-ship calls.

This is not a role of running an existing validation harness and reporting its output. When a parity check fails, you need to identify the operation responsible for the divergence and explain why it diverges. The tooling speeds up that investigation but does not replace it.

Responsibilities

• Decide what equivalence means for each optimized operation relative to its reference, then design inputs that would reveal a violation if one existed.
• Trace low-level numerical differences from quantization, compilation and reduced precision to driving behavior, measured with open-loop metrics such as trajectory displacement error and perception IoU, and closed-loop outcomes. Explain the mechanism behind each link, not only the correlation.
• Apply Jacobian and Hessian methods to measure how outputs respond to weight and input perturbations, extend this to out-of-distribution inputs, and package it as tooling that engineers other than you can run on every checkpoint.
• Design diagnostics that catch and explain training instability, including gradient vanishing, gradient explosion, loss spikes and silent divergence. Split gradient and update trajectories into loss-descent and oscillatory parts, including under schedules such as WSD.
• Keep metrics, gradient decomposition and diagnostic logging cheap enough to run inside production distributed training, with almost no throughput cost and no risk of running out of memory. Observability that teams cannot afford to switch on does not help.

Requirements

• Strict requirement: a working command of numerical analysis and matrix theory, including Jacobian and Hessian estimation, spectral properties, conditioning and floating-point error. You should know why an estimator's variance can blow up and when that matters.
• Strict requirement: a solid model of how neural networks train, covering loss landscapes, gradient and error propagation, optimizer behavior and the ways training fails. You should have a view on what a gradient norm spike does and does not indicate.
• Strict requirement: a habit of probing numerical edge cases such as denormals, extreme dynamic range, catastrophic cancellation, degenerate shapes and accumulation order. You should ask first what a passing test failed to cover.
• Strict requirement: high proficiency in PyTorch and Python, with a record of building analytical tools that are mathematically sound and fast enough for production training and evaluation.
• Strict requirement: the ability to turn a numerical result into something an engineer who skips your derivation can act on, including picking the quantity that answers the question and tracing an odd number back to its source operation.
• Strict requirement: a Bachelor's, Master's or PhD in Applied Mathematics, Control, Physics, Computer Science, Data Science or a closely related quantitative field.
• Not required on day one, but expected to be learned over three to nine months: the training recipes and acceleration techniques used in the stack; the semantics of autonomous driving behavioral metrics and what counts as meaningful behavioral drift for a vehicle; the model compression portfolio, deployment path to the car and organizational context around ship decisions; and the team's parity validation workflow, including where the quantization and compilation toolchains cause problems.
• Useful but not required: experience debugging large training runs, such as loss spikes and numerical divergence; distributed training beyond DDP, such as FSDP, Megatron-LM, DeepSpeed or 3D parallelism, especially low-overhead observability over sharded parameters and gradients; quantization, compiler toolchains or inference-time numerical parity; building evaluation pipelines and metrics for AV or ADAS systems; published or applied work on robustness, OOD generalization or adversarial and perturbation analysis.

Conditions

Base salary range: $170,600 to $261,300. Actual pay within the range depends on job-related factors. This is a good faith estimate under applicable state law and may not reflect roles outside New York, Colorado, California or Washington.

An incentive bonus program pays out based on company performance, job level and individual performance.

Hybrid role: you are expected to be on site at least 3 days a week, or as your manager sets.

Travel: up to 25% is required.

Relocation assistance is available for eligible candidates under company policy.

Equal opportunity: the company does not tolerate unlawful discrimination and works to build an inclusive workplace.

Benefits

• Medical, dental and vision coverage, a Health Savings Account, Flexible Spending Accounts, a retirement savings plan, sickness and accident benefits, life insurance, and paid vacation and holidays.
• Well-being and career support available from the first day, described in the company's Total Rewards resources.