Senior Agent Evaluation Engineer, AI Evaluation Platform
Confirmed on the employer's own hiring board on Oct 30, 2026. First seen by Alion on May 19, 2026. Apple scores A on the Alion truth index.
Join Apple Services Engineering to define how we measure whether AI agents actually work. We are seeking a senior engineer who has built agents and knows how to evaluate them rigorously. You will design evaluation suites, build and validate judges, and ship agents on our evaluation platform, so the platform is shaped by real use and not guesswork. You will partner with researchers, product teams, and platform engineers to turn evaluation methodology into tools other teams can adopt. We are looking for builders who move quickly, iterate in the open, and can tell a trustworthy evaluation from a theatrical one.
Description
We're building the evaluation platform that will serve all of Apple's generative AI and agent systems. Building agents got faster. Knowing whether they got better did not. This role closes that gap.
You'll build real agents and evaluate them: multi-turn trajectories, tool-use chains, non-deterministic outputs, and judge reliability. You'll write evaluation suites, validate judges against human labels, and measure variance across trials. Some agents you build will evaluate other systems (test-case generation, judging, failure triage). Others will be reference agents or built with partner teams. When the platform gets in your way, you'll file the gap, and often fix it yourself.
This is a hands-on role with a lot of autonomy. You'll work in Python with a team of 7-12 engineers.
Minimum Qualifications
* 4-8 years of software engineering experience shipping production software.
* Hands-on agent builder. You've built LLM agents beyond demos: tool use, MCP, multi-step workflows, context management, error recovery.
* Evaluation depth. You've designed evaluations for agents, not just models. You understand non-deterministic outputs, multi-trial measurement (pass@k, variance), trajectory scoring, and scoring drift. You've worked with tools like Inspect, Braintrust, LangSmith, or Promptfoo, or built your own.
Judge rigor. You know how to build an LLM judge, validate it against human agreement, and know when not to trust it.
Fast from idea to prototype. You can point to things you built quickly, then iterated on or shipped. You can also tell us about a prototype you killed, and why.
Fluent with AI coding tools. You use Claude Code or equivalents daily, including skills, slash commands, and agent-style workflows.
Python as primary language. Strong with FastAPI, Pydantic, and the ecosystem. Your code is clean, tested, and easy for the next person to pick up.
Judgment on speed vs. discipline. You know when a quick prototype is enough and when you need a design doc, tests, and a careful rollout.
Cross-team collaboration. You've worked closely with researchers, product teams, or platform engineers. You write clearly and flag blockers early.
Preferred Qualifications
Experience working on developer platforms, internal tools, or SDKs
Production experience with LLM/agent systems - building, evaluating, or operating them
Familiarity with job orchestration frameworks (Temporal.io, Airflow, or similar)
Distributed compute experience (Ray, Dask, or Kubernetes-based job systems)
Experience with experiment tracking or ML lifecycle tooling (Weights & Biases, MLflow, etc.)
Startup or early-stage experience where you wore multiple hats and shipped under constraint