Most model evaluation stops at a scorecard. The failing traces get thrown away, even though they're the most relevant training signal we have. Our team puts them to work. We evaluate across offline benchmarks, device-in-the-loop simulation, and on-device observation in production, and we turn what we find into persona-grounded datasets and reward signals for foundation model post-training.

In this role, you will build the systems that connect failure analysis to measurable model and agent improvements across a broad range of AI products at Apple. The path from finding a problem to fixing it is short: the loop you build finds the failure, produces the data and checks the result. The work reaches millions of users.

Minimum Qualifications

Strong background in machine learning and distributed systems.
A proven track record of solving ambiguous problems end to end: defining the problem and the metric, building the solution, and showing measured impact.
Experience leading technical work across multiple codebases, teams, and organizations.
Experience building and maintaining ML infrastructure for evaluation, training, or deployment.
6+ years of professional experience as a software engineer, preferably in machine learning or a related field. Bachelor's or Master's degree in Computer Science or a related field, or equivalent practical experience.

Preferred Qualifications

Experience with LLM evaluation, LLM-as-judge, or reward modeling.
Experience with prompt optimization, agent harness development, or post-training (SFT, DPO, RLHF).
Proficiency in Python and ML frameworks such as PyTorch or JAX.
Experience with agentic systems, simulation environments, or trajectory-based data generation.
Familiarity with on-device or privacy-preserving ML.
Understanding of how training data is used in post-training and how data choices affect model behavior.
Excellent communication skills.