Senior Ml Applied Scientist , Frontier Ai Startup Team
Amazon Web Services, Inc.
Seattle
Full-time
Senior-level
AI-native startups now train their own models, post-train open-weight models for their products, and serve inference at volumes where cost per output decides their margins. The startup team of the AWS Frontier AI Services puts applied scientists to work directly with those founders and their engineers. We're hiring a Senior Applied Scientist in North America to own those engagements with our strategists end to end: scope the problem with the founding team, design and run the experiments, and ship working code into the startup's stack on AWS. The problems are open-ended and the timelines run in weeks, so your scientific judgment sets both the approach and the bar for done. Key job responsibilities - Work directly with founders, CTOs and ML leads at AI startups to define the problem and agree on a production target and an evaluation plan. Then deliver against them in six-to-twelve-week engagements, with the startup's own engineers working alongside you. - Improve model quality for the startup's use case by post-training open-weight models. That covers supervised fine-tuning, preference optimization, reinforcement learning (for example GRPO), distillation, and the synthetic data and evaluation sets that show whether it worked. - Cut training and inference cost and latency on NVIDIA GPUs and AWS Trainium. Profile where the compute goes, apply quantization, speculative decoding, KV-cache reuse, batching, parallelism and kernel-level tuning, and measure cost per output before and after. - Invent new methods when published ones fall short of the customer's target, and publish the results at peer-reviewed venues, on AWS blogs and as open-source code. - Turn each engagement into assets the next one can reuse: reference implementations, evaluation harnesses and playbooks. Take the gaps you find in Amazon Bedrock, Amazon SageMaker and AWS Trainium back to those service teams as specific requests. - Raise the science bar on the team. Review experiment designs and code, mentor scientists and engineers, and help account teams and strategists judge which startup requests need a scientist. A day in the life You start the day on a call with the CTO of a robotics startup whose vision-language model misses objects in low light. Before anyone changes the model, you agree on an evaluation set built from their own camera footage. Mid-morning you check an overnight reinforcement-learning run for another startup: reward is up but two task categories regressed, so you write up why and what to try next. After lunch you join an AWS account manager and a GenAI strategist to scope a request from a voice AI company that needs lower inference latency before it scales up. You leave with a two-week profiling plan and one go/no-go metric. You end the day reviewing a teammate's experiment design and editing a workshop paper drawn from last quarter's engagement. About the team We're applied scientists and GenAI strategists working as one team with venture-backed, AI-native startups, from the first discovery call to a model running in production. The work spans large language and vision-language models, agents, voice, multimodal retrieval and physical AI, including world and action models. We judge an engagement by what the startup ships. Startups make platform decisions in weeks, so we scope quickly, share results early, and say plainly when something isn't working. This role focuses on startups in North America. You'll work with AWS account teams, solutions architects and service teams, and with Amazon's wider science community.