广告全漏斗智能代理软件开发高级经理

FAIM Evaluations.

Amazon Advertising is building toward a future where an advertiser specifies a few marketing parameters (budget, success definition, which products to promote) and a set of AI agents handles the rest.

The Full Funnel Agentic Intelligence and Models (FAIM) organization owns that bet: the Ads Nova agent, the Ads Nova model it reasons with, and the agent infrastructure, learning environments, and evaluations that connect the two.

FAIM Evaluations team.

You will report directly to the Vice President of Full Funnel Agentic Intelligence and Models and own how the entire organization answers one question: is this actually good at advertising?

This is a ground-up build.

Evaluation today lives inside individual model and agent teams, measured task by task.

You will create the standalone engineering team that turns it into a shared, rigorous system spanning the Ads Nova model, the Ads Nova agent, and the internal agent that serves our Sales, Services, and Operations teams.

No inherited harness, no pattern to follow, and a direct line to the VP who sponsors the work.

What we're building An advertising benchmark: a representative set of real advertising tasks, organized by domain and difficulty, from single-step questions through multi-step analysis to long-horizon strategic work, each with structured criteria for what a correct end-to-end response looks like Evaluation infrastructure that scores models and agents deterministically against that benchmark, compares Ads Nova to frontier models on the tasks that matter to advertisers, and gives every science and product team in

FAIM the same yardstick Rubrics and task environments built to serve double duty: scoring quality today and producing the reward signal that trains the next version of the model An expert-in-the-loop program that captures how experienced advertising practitioners actually work and encodes that judgment into criteria a machine can grade against A capability map, derived from benchmark results, that tells

FAIM where the model and agent stand and what to train next Key job responsibilities Found and lead a standalone team of roughly 10-12: software engineers plus a product manager and a technical program manager;

FAIM evaluations end to end: task taxonomy, rubric design, environment construction, scoring, benchmark versioning, and the separation between what we evaluate on and what we train on Build for the whole org, not one product: your team evaluates the Ads Nova model, the Ads Nova agent, and the internal agent, and you participate in the planning and reviews for all three Partner with applied scientists across

FAIM to turn evaluation criteria into training signal for reinforcement learning, and to make sure what we measure is what we optimize Run the domain-expert program: source advertising practitioners, define the annotation and calibration process, and hold the quality bar on inter-rater agreement Set the evaluation standard for the organization and hold the line on it;

where good internal assets already exist, adopt them rather than rebuild Publish results leadership and partner teams trust, and own the cadence for re-scoring as models, agents, and tasks evolve Represent evaluations in VP-level reviews, annual planning, and cross-org discussions on model and agent quality We're looking for a leader who brings 10+ years of engineering experience and 5+ years managing engineering teams, including building a team from a small core A track record delivering evaluation systems, benchmarks, or data-quality programs for machine learning models, ideally large language models or agentic systems Working fluency in how modern models are trained and improved (supervised fine-tuning, reinforcement learning from rubric or verifier signal) and what makes an eval useful as a training asset rather than only a scorecard Judgment about measurement: when all-or-nothing grading beats partial credit, how to find ambiguous criteria through grader disagreement Experience running expert-annotation or labeling programs with external partners, including quality control at scale The ability to operate in ambiguity: turn "is it good at advertising" into a concrete, scored, versioned asset with minimal scoping help Comfort working as the engineering counterpart to scientists you do not manage, and the influence to get model, agent, and product teams onto one yardstick Advertising domain knowledge, or the curiosity and speed to build it by working closely with practitioners Experience with Amazon Bedrock, agent frameworks, and tool-use protocols (MCP) is a plus

10+ years of engineering experience and 5+ years managing engineering teams, including building a team from a small core.

A track record delivering evaluation systems, benchmarks, or data-quality programs for machine learning models, ideally large language models or agentic systems.