Amazon Devices is an inventive research and development company that designs and engineers high-profile devices like the Kindle family of products, Fire Tablets, Fire TV, Health & Wellness, Amazon Echo, and Astro products.

This is an exciting opportunity to bring generative AI to Amazon's consumer products, both on-device at the edge and in the cloud. Our compression platform delivers 20x to 100x neural network compression, but using it well still takes weeks of hands-on learning and expert intuition. The Edge AI Model Studio team exists to change that. We become the expert users so partner teams don't have to: we turn compression science into reliable, production workflows, and we package the results into a library of compression-ready student architectures that partners can run on their own. Our north star is simple. Training-to-deployment should feel like pushing a button, not a month-long science project.

We are looking for an Applied Scientist to join Model Studio and help compress the next generation of models for edge and cloud deployment across modalities, including large language models, vision-language models, speech and audio models, and omni models that reason jointly over text, audio, and video. You will apply and extend state-of-the-art compression recipes to real models, define the benchmarks and evaluation methodology that make trade-offs explicit, and build the reference implementations that let other teams deploy compressed models without our help. You will work backwards from deployment constraints such as memory, latency, throughput, power, and cost, which differ across edge and cloud targets, partnering closely with fellow scientists, platform and compiler engineers, hardware architects, and product teams. The role sits on two frontiers at once. Compressing a model effectively and healing it back to quality means staying current not just with the latest compression techniques, but with the rapidly evolving model architectures themselves, and understanding deeply how each one works inside.

You will take ownership of project-level delivery, apply advanced compression across a wide range of real models, and have room to grow your scope and technical influence.

Key job responsibilities

• Apply and extend compression recipes (knowledge distillation, structured pruning, and post-training and quantization-aware quantization including low-bit and mixed-precision) to assigned models, achieving 20x to 100x compression while preserving model quality.
• Design and run healing recipes (fine-tuning and distillation that recover accuracy lost to compression), iterating on data mixes, objectives, and training settings until the compressed model meets its quality bar.
• Track emerging model architectures and dissect how they work internally, so you can choose where to compress, anticipate where accuracy will break, and design recovery strategies grounded in the model's actual structure.
• Build a library of compression-ready model entries: reference implementations, compression recipes, model cards, and benchmark results that partner teams can run self-service to produce deployment-ready artifacts for edge and cloud targets.
• Define the datasets, benchmarks, and KPIs that matter for your models, and build evaluation methodology that makes accuracy, latency, memory, and cost trade-offs explicit.
• Run fast feasibility gates on new model families and modalities before committing to long efforts, and pivot early when a candidate does not clear the bar.
• Capture platform friction as high-signal feedback: minimal reproductions and tracked fix requests that help platform and compression-science partners root-cause issues, so partner teams never rediscover the same blockers.
• Write reproducible, testable, well-documented code that meets the SDE I bar, so your recipes and results can be reproduced and built on by others.
• Collaborate with Applied Scientists, platform and compiler engineers, hardware architects, and partner teams; mentor interns and help newer teammates ramp up.
• Where appropriate and not precluded by business considerations, publish and present on Amazon's behalf at top ML venues such as NeurIPS, ICLR, and MLSys.

A day in the life
You pick up a vision-language model whose vision tower needs to fit tight memory, latency, and cost budgets for deployment. You configure a quantization-aware training run at the team's target compression ratio, then check the compressed checkpoint against a visual reasoning benchmark and find it recovers only part of the baseline accuracy. You design a healing run to close the gap, tuning the data mix and training objective to fine-tune the compressed model back toward the teacher's quality. The next checkpoint clears most of the gap but still lands short, so rather than assume the recipe is at fault, you dig into the evaluation harness and discover a benchmark filter is misaligned, deflating the score. You fix the filter, re-run, and confirm the healed model lands where the science predicts. You then package the work as a reusable model entry (recipe, model card, benchmark numbers, and a reference implementation a partner team can run on their own) and file a minimal reproduction of the harness bug so no one rediscovers it.

A typical week mixes hands-on compression and evaluation with design discussions alongside fellow scientists and platform engineers. You run a fast feasibility gate on a new model family before committing to a long effort, profile a compressed model to confirm a real throughput gain, and turn a recurring friction point into a reusable pattern. You work in a small, fast-moving team where every recipe you harden compounds across future models and every partner you unblock ships faster.

About the team
We compress frontier models 20x to 100x and put them in the hands of millions of customers, everywhere from your pocket to the cloud: the device in your hand, the Echo on your counter, and the services behind them. The models the industry shipped last month, we are shrinking this month, across language, vision, speech, and omni. That is the job: take the best models in the world and make them small enough, fast enough, and cheap enough to run everywhere, without giving up the intelligence that makes them worth running.

Edge AI Model Studio is the team that makes it real. We are the expert users of a compression platform that most of Amazon cannot yet wield, and our mission is to change that, turning weeks of expert intuition into recipes anyone can run. We are small, we move fast, and we own our work end to end: a result counts only when it ships with a recipe, benchmarks, and an artifact a partner team can run without us. Every recipe we crack compounds across every model that follows. If you want your science in real products at real scale, and you want to put the frontier of generative AI in the hands of millions of customers, come build it with us.