Principal Applied Scientist, Mission AI Lab, AWS Region Services

Some of the hardest problems in security are waiting on models that do not exist yet. As a Principal Applied Scientist, you will create them. You start from the strongest models in the world and push them to excel at problems no off-the-shelf model has ever seen. This is one of the largest AI infrastructure programs AWS has ever taken on, and you will train on latest-generation GPU hardware, among the first anywhere to put it to work at this scale.

Much of this has no established playbook. There is no benchmark to tell you a model is ready, so you will invent both the method that gets it there and the measure that proves it. The lab is purpose-built for this work and you join at the start, helping shape its technical direction and building the methods and evaluations that carry forward into everything it does next.

This is a hands-on role. You will research the state of the art science, make the calls on how models are trained and evaluated, write code and work alongside the engineers who run the infrastructure, and partner directly with the analysts and operators who depend on the work. You will also explain what the models can and cannot yet do, in plain terms, to the people whose decisions rely on them.

The candidate selected must obtain and maintain a security clearance at the TS/SCI with polygraph level. Upon start, the selected candidate will be sponsored for a commensurate clearance for each government agency for which they perform AWS work.

Key job responsibilities

• Own the end-to-end adaptation of frontier models for complex reasoning tasks, from supervised fine-tuning and post-training through reinforcement learning, delivered as production systems people can rely on in the field
• Define what "good" means for each task, building the evaluation sets, test harnesses, and metrics that measure model quality where no public benchmark exists
• Build and own the data pipelines that turn raw operational data into high-signal training and evaluation sets, including curation, quality scoring, provenance tracking, and strict leakage and contamination controls
• Judge when a problem needs a genuinely new method versus an extension of an existing one, and make the call with evidence
• Choose the training and adaptation methods, including parameter-efficient techniques, that turn serious dedicated compute into the largest gains in capability
• Extend adapted models into new priority domains, owning transferability, generalization, and calibration under real-world distribution shift
• Run data-preparation and human-in-the-loop feedback loops with the analysts who know the problems best, turning their expertise into better data and sharper evaluations
• Package methods and evaluations so they can be reused as the lab"s work grows, and shared with Amazon"s model teams
• Represent the lab"s science in technical reviews and to program and technical leaders, surfacing what the models can do and where the risks are
• Help set the technical bar for the lab, and mentor scientists as the team grows

A day in the life

You start the morning reviewing an overnight fine-tuning run against an evaluation suite you built, checking whether the adapted model beats the base on the metrics that matter most. Mid-morning, you work with the engineers who run the infrastructure to get a large training run moving and squeeze the most capability out of it. After lunch, you sit with domain analysts to turn their feedback into a cleaner, better-labeled evaluation set for a change-detection task, then decide whether the next gain calls for a new method or an extension of one you already have. You close the day documenting that method and evaluation so the rest of the lab can build on it, and writing a short, plain-language readout of what the model can and cannot yet do for the people who depend on it.

About the team

The Mission AI Model Lab is the science team behind AWS"s Mission AI program. We take existing models and make them useful for problems that matter to national security, working on dedicated AI infrastructure built for exactly that. This lab is purpose-built and these are among its first scientists, so the people who join now help shape how it works and what it builds. We sit where frontier-model research meets national security delivery, and we value invention on genuinely new problems, evaluation we can defend, and plain talk with the people who rely on what we build.