Senior Applied Scientist, Frontier AI Assets - Assessments (FAI)

Job description

Amazon's Artificial General Intelligence (AGI) organization builds frontier models and the AI agents on top of them, and every one of them depends on data we can trust. The Frontier AI (FAI) Assessment team owns the science of dataset quality — for the data that trains frontier models, and for the benchmarks that determine whether a model or an agent actually works.

In this role you will build the automated systems that make quality assessment at scale. Large language model (LLM)-as-a-Judge is the starting point, and our goal is to develop an agentic system that plan its own audits, critiques its own judgments, and improves its own accuracy over time. You will design it, calibrate it against expert human judgment, and set the technical direction for how the organization measures data quality.

Key job responsibilities

- Design the assessment methodology for frontier AI assets

- Build the automated assessment system, from LLM-as-a-Judge baselines to agents that plan their own audits and learn from their own errors

- Oversee the expert audit program and raise the proficiency of the auditors who run it

- Design the measurement methods that make quality findings defensible, including sampling strategy and error analysis

- Communicate findings to the teams that create the data and to the teams that train models and build agents using it

- Set technical direction for assessment science, mentor junior scientists, and present results to senior leadership

- Publish research on assessment methodology at top-tier venues and contribute to patents

A day in the life

- Train AI agents to assess the quality of datasets and evaluation benchmarks, and document where they fall short

- Diagnose their failures and improve the models, rubrics, and calibration behind them

- Review the quality reports that auditors and agents produce, and decide whether the conclusions hold

- Coach expert auditors and junior scientists, and move more of the manual audit work into automation

- Lead research on self-improving assessment agents, from open problem to publication

About the team

FAI Assessment is part of Frontier AI Assets in AGI. We assess the quality of the datasets and benchmarks behind Amazon's frontier models and agents, and we define what good data means. Today that work relies on human-in-the-loop review by domain experts. Our aim is to automate it with self-improving agents that experts keep calibrated. We are a small team of scientists working closely with the data, modeling, and agent teams.