Fitzgerald keynote.png
Amazon senior applied scientist Jack FitzGerald, delivering a keynote talk at the joint Language Intelligence @ Work and SEMANTiCS conference in Vienna, Austria.

Scaling multilingual virtual assistants to 1,000 languages

Self-supervised training, distributed training, and knowledge distillation have delivered remarkable results, but they’re just the tip of the iceberg.

Yesterday at the joint Language Intelligence @ Work and SEMANTiCS conference in Vienna, Austria, Amazon senior applied scientist Jack FitzGerald delivered a keynote talk on multilingual virtual assistants and the path toward a massively multilingual future. This is an edited version of his talk.

The evolution of human-computer interaction paradigms

In the past 50 years, computing technology has progressed from text-based terminal inputs, to graphical user interfaces, to predominantly web-based applications, through the mobile era, and finally into the era of a voice user interface and ambient computing.

Interface timeline.png
A brief history of computing interfaces.

Each of these paradigms has its own challenges with respect to multilingualism, whether it was the migration from ASCII to Unicode or proper character rendering on a website. However, I would argue that a voice AI system is the most difficult paradigm yet with respect to massive multilingualism.

The first reason is that the input space for voice interface commands is unbounded: the user can phrase each command in hundreds of different ways, all of which are valid. Another reason is that even within a single language, there can be many different dialects and accents.

Related content
Amazon Visiting Academic Barbara Poblete helps to build safer, more-diverse online communities — and to aid disaster response.

Most important, the coupling between language and culture is inescapable. Whether it’s the level of formality used, preferred activities, or religious differences, there isn’t a one-size-fits-all solution. Instead, we must adapt the virtual assistant to understand cultural context and say only things that are appropriate for a given locale.

Voice AI systems today

A typical voice AI system includes automatic-speech-recognition models, which convert raw audio into text; natural-language understanding models, which determine the user’s intent and recognize named entities; a central service for arbitration and dialogue management, which routes commands to the proper services or skills; and finally, a text-to-speech model, which issues the output. Additional tasks might include expansion of the underlying knowledge graph and semantic parsing, localization of touch screen content, or local information services.

Alexa overview.png
An overview of Alexa’s design.

Let’s look at some of the operational considerations for supporting multiple languages in such models. One is the training data: they must be topically exhaustive, meaning that they cover the full spectrum of possible user utterances, and they must be culturally exhaustive — for instance, covering all of the holidays a user might celebrate. They must also remain up-to-date, and it’s not always easy to add something new to the model without regression on existing functionalities.

A second consideration is in-house testing. Though in many cases one can get away with synthetic or otherwise artificial data for model training, for testing it’s important to have realistic utterances. Those typically need to come from humans, and collecting them can be a major expense. It’s also useful to perform live, interactive testing, which requires people who can speak and understand each language that the system supports.

Related content
New approach corrects for cases when average improvements are accompanied by specific regressions.

Finally, it’s important to have the ability to support users and process their feedback. In most cases, this again requires staff who understand each of the supported languages.

Ultimately, human-based processes are not very scalable if our goal is to support thousands of languages. Instead, we must turn to technology to the greatest extent possible.

Multilingual modeling today

One of the leading reasons for the current success of multilingual text models is self-supervision.

In traditional supervised learning, a model would be trained from scratch on the desired task. If we wanted a model that would classify the sentiment of a product review, for example, we would manually annotate a bunch of product reviews, and we would use that dataset to train the model.

Today, however, we make use of transfer learning, in which text models are pretrained on terabytes of text data that don’t require manual annotation. Instead, the training procedure leverages the structure inherent to the text itself.

Self-supervision signals.png
Self-supervised-training objectives.

We’ll call this self-supervised pretraining With the masked-language-modeling training objective, for instance, the model is fed the input “for [MASK] out loud!”, and it must predict that “[MASK]” should be filled with the word “crying”. Other objectives, such as causal language modeling, span filling, deshuffling, and denoising can also be used.

Because the datasets required for self-supervised pretraining are unlabeled and monolingual, we can leverage troves of data, such as Common Crawl web scrapes, every Wikipedia page in existence, thousands of books and news articles, and more. Couple these large datasets with highly parallelizable architectures such as transformers, which can be trained on over a thousand GPUs with near linear scaling, and we can build models with tens or hundreds of billions of dense parameters. Such has been the focus for many people in the field for the past few years, including the Alexa Teacher Model team.

One incredible consequence of the transfer learning paradigm is called zero-shot learning. In the context of multilingual modeling, it works like this: the modeler begins by pretraining the model on some set of languages, using self-supervision. As an example, suppose that the modeler trains a model on English, French, and Japanese using every Wikipedia article in those three languages.

Related content
New end-to-end approach to zero-shot video classification dramatically outperforms predecessors.

The next step is to adapt the model to a particular task using labeled data. Suppose that the modeler has a labeled dataset for intent classification, but only in English. The modeler can go ahead and fine-tune the model on the English data, then run it on the remaining languages.

Despite the fact that the model was never trained to do intent classification with French or Japanese data, it can still classify intents in those languages, by leveraging what it learned about those languages during pretraining. Given that the acquisition of labeled data is often a bottleneck, this property of language models is highly valuable for language expansion. Of course, zero-shot learning is just the extreme end of a continuum: transfer learning helps even out performance when the labeled data in different languages is imbalanced.

Zero-shot multilingual.png
Zero-shot learning for multilingual adaptation.

The next step up the data efficiency ladder is performing tasks without any additional training or fine tuning, using only a couple of labeled records or none at all. This is possible through “in-context learning,” which was popularized in the GPT-3 paper.

To perform in-context learning, simply take a pretrained model and feed it the appropriate prompts. Think of a prompt is a hint to the model about the task it should perform. Suppose that we want the model to summarize a passage. We might prefix the passage with the word “Passage” and a colon and follow it with the word “Summary” and a colon. The model would then generate a summary of the passage.

Related content
In the past few years, advances in artificial intelligence have captured our imaginations and led to the widespread use of voice services on our phones and in our homes.

This is the zero-shot in-context learning case, meaning that no fine-tuning is performed, and no labeled data are needed. To improve task performance, we can feed a few examples to the model before asking it to perform the task. Though this does require some labeled data, the amount is small, usually in the tens of examples only.

Our Alexa Teacher Model team recently trained and tested a 20-billion-parameter sequence-to-sequence model that was multilingual and showed nice performance for in-context learning. For example, we showed state-of-the-art performance on machine translation with in-context learning. The model can achieve competitive BLEU scores even for some low-resource languages, which is incredible given that no parallel data was used during pretraining, and no labeled data besides a single example was used at any step in the process.

We were particularly proud of the relatively small size of this model, which could compete with much larger models because it was trained on more data. (The Chinchilla model from OpenAI showed a similar result.) Though a large model trained on a smaller dataset and a smaller model trained on a larger dataset may use the same total compute at training time, the smaller model will require less compute and memory during inference, which is a key factor in real applications.

Given that models demonstrate multilingual understanding even without labeled data or parallel data, you may be wondering what’s happening inside of the model. Since the days of word2vec and earlier, we’ve represented characters, words, sentences, documents, and other inputs as vectors of floats, also known as embeddings, hidden states, and representations. Concepts cluster in certain areas of the representational space.

Related content
Training a product discovery system on many languages at once improves performance in all of them.

As humans, we can think only in three dimensions, whereas these representations are high-dimensional, but you can visualize this clustering in two or three dimensions as a reductive approximation. All the languages the model supports would cluster the concept of sitting in a chair in one region of the representational space; the concept of the ocean would inhabit a different cluster; and so forth.

Indeed, Pires et al. have shown that synonymous words across languages cluster together in the mBERT model. When examining 5,000 sentence pairs from the WMT16 dataset, they found that, given a sentence and its embedding in one language, the correct translation from another language is the closest embedding to the source embedding up to 75% of the time.

This manner of clustering can also be manipulated by changing the objective function. In their work on speech-to-text-modeling, Adams et al., from Johns Hopkins, were seeing undesirable clustering by language, rather than by phonemes, in the representational space. They were able to correct by adding training objectives around phoneme prediction and language identification.

The Alexa Teacher Model distillation pipeline

Once we have multilingual models, how do we adapt them to a real system? At the recent KDD conference, we presented a paper describing the Alexa Teacher Model pipeline, consisting of the following steps.

First, a multilingual model with billions of parameters is trained on up to a trillion tokens taken from Common Crawl web scrapes, Wikipedia articles, and more. Second, the models are further trained on in-domain, unlabeled data from a real system. Third, the model is distilled into smaller sizes that can be used in production. The final models can then be fine-tuned using labeled data and deployed.

ATM pipeline.png
The Alexa Teacher Model (AlexaTM) pipeline. The Alexa Teacher Model is trained on a large set of GPUs (left), then distilled into smaller variants (center), whose size depends on their uses. The end user adapts a distilled model to its particular use by fine-tuning it on in-domain data (right).

In tests, we found that our model was more accurate than a publicly available pretrained model fine-tuned on labeled data, and it significantly reduced customer dissatisfaction relative to a model trained by a smaller teacher model (85 million parameters, say, instead of billions). In short, we’ve verified that we can leverage the additional learning capacity of large, multilingual models for production systems requiring low latency and low memory consumption.

Scaling to 1,000 languages

I mentioned the fascinating ability of language models to learn joint representations of multiple languages without labeled or parallel data. This ability is crucial for us to scale to many languages. However, as we scale, we need test data that we can trust so that we can evaluate our progress.

Related content
MASSIVE dataset and Massively Multilingual NLU (MMNLU-22) competition and workshop will help researchers scale natural-language-understanding technology to every language on Earth.

Toward this end, my team at Amazon recently released a new benchmark for multilingual natural-language understanding called MASSIVE, which is composed of one million labeled records spanning 51 languages, 18 domains, 60 intents, and 55 slots. All of the data were created by native speakers of the languages. We also released a GitHub repository with code that can be used as a baseline for creating multilingual NLU models, as well as leaderboards on eval.ai.

Now, you may retort that 51 languages is still a long ways from 1,000 languages. This is true, but we purposefully chose our languages in order to maximize typological diversity while staying within our budget. Our languages span 29 language genera, 14 language families, and 21 distinct scripts or alphabets. The diversity of the chosen languages allows a modeler to test technology that should scale to many more languages within each represented genus, family, and script.

That said, we certainly have some major gaps in language coverage, including across native North and South American languages, African languages, and Australian languages. Yet we are optimistic that our fellow researchers across the field will continue to produce new labeled benchmark datasets for the world’s thousands of low-resource languages.

Massive languages.cropped.png
The 51 languages of MASSIVE, including scripts and genera.

Another difficulty with our current modeling approaches is that they rely on data sources such as web scrapes, encyclopedic articles, and news articles, which are highly skewed toward a small set of languages. Wang, Ruder, and Neubig recently presented some fascinating work leveraging bilingual lexicons — corpora consisting of word-level translations — to improve language model performance for low-resource languages. Lexicons cover a far greater portion of the world’s languages than our typical data sources for language modeling, making this an exciting approach.

Related content
Self-learning system uses customers’ rephrased requests as implicit error signals.

Researchers, missionaries, and businesspeople have been created fundamental linguistic resources for decades, from Bible translations to the Unimorph corpus. The Unimorph datasets are used for the SIGMORPHON shared task, in which a model must predict the correct formulation of word given that word’s root and certain morphological transformations, such as part of speech, tense, and person. We must find more ways to leverage such resources when creating massively multilingual voice AI systems.

As a final technique for scaling to many more languages, we can consider what we in Alexa call “self-learning.” Some of my Alexa colleagues published a paper showing that we can mine past utterances to improve overall system performance. For example, if a user rephrases a request as part of a multiturn interaction, as shown on the left in the figure below, or if different users provide variations for the same desired goal, as shown on the right, then we can make soft assumptions that the different formulations are synonymous.

All of these cases can be statistically aggregated to form new training sets to update the system, without the need to manually annotate utterances. In a multilingual system, such technology is particularly valuable after the initial launch of a language, both to improve performance generally and to adapt to changes in the lexicon.

Self-learning.png
Alexa’s self-learning mechanism.

The road ahead

I hope that you share my wonder at the current state of the art — the scale of language-model training, the magic of zero-shot learning, and the distillation of knowledge into compact models that can run in latency-sensitive systems. All of this is incredible, but we’ve only scratched the surface of supporting the world’s 7,000 languages.

To move into the next era of massive multilingualism, we must build new and increasingly powerful models that can take advantage of low-cost data, particularly unlabeled monolingual data. We must also build models that can leverage existing and upcoming linguistic resources, such as bilingual lexicons and morphological-transformation databases. And finally, we must expand available language resources across more languages and domains, including more unlabeled monolingual corpora, more parallel resources, and more realistic, labeled, task-specific datasets.

Increased multilingualism is a win for all people everywhere. Each language provides a unique perspective on the world in which we live. A rich plurality of perspectives leads to a deeper understanding of our fellow people and of all creation.

Keep building.

Research areas

Related content

US, CA, Pasadena
The Amazon Web Services (AWS) Center for Quantum Computing (CQC) in Pasadena, CA, is hiring an Applied Scientist to lead integration research and development for superconducting quantum devices. You will design, model, and test the chip-to-chip, chip-to-package, and signal-delivery interfaces that let superconducting quantum computers scale, and you will characterize integrated assemblies at cryogenic temperatures to learn how integration choices affect qubit performance, yield, and manufacturability. You will join a multi-disciplinary team of theoretical and experimental physicists, materials scientists, and hardware and software engineers working at the forefront of quantum computing. You will carry integration technologies from concept through hardware demonstration, build the test standards and automation that make results reproducible across devices, and collaborate with fabrication, hardware, and circuit design engineers to remove the dominant sources of yield loss. You will also contribute to multi-year roadmap planning for integration and scaling capabilities and help establish the scientific foundation for this capability area within the Device team. Key job responsibilities In this role you will develop the integration technologies that let superconducting quantum computers scale. Your work will span four connected areas: - Integration research and development: Design, model, and test integration solutions for superconducting quantum devices, including chip-to-chip, chip-to-package, and signal-delivery interfaces. Build electromagnetic, thermal, and mechanical models to predict and optimize performance, and characterize integrated assemblies at cryogenic temperatures. - Scaling technologies: Develop and evaluate approaches to scaling superconducting quantum hardware against yield, reusability, manufacturability, and scalability, and carry the most promising ones from concept through hardware demonstration. - Test standards and automation: Define test standards that give rigorous, reproducible measurements of how integration affects qubit coherence, loss, and crosstalk. Automate data collection and reporting against these standards so characterization scales across devices and configurations. - Yield and manufacturability: Work with fabrication, hardware, printed circuit board, and circuit design engineers to identify and remove the largest sources of yield loss and integration uncertainty, and to design processes and components that are reused across device generations and manufactured at scale. About the team The Device team at the AWS Center for Quantum Computing designs, builds, and measures the superconducting quantum devices at the heart of our processors, and develops the integration and packaging that turn individual devices into working systems. You will work alongside physicists, materials scientists, and hardware and software engineers across design, fabrication, measurement, and theory, in a lab built for rapid iteration from concept to cryogenic test. Over the coming years the team's focus is making superconducting quantum hardware scale, and this role establishes integration as a scientific capability area within that effort. Why AWS? Amazon Web Services (AWS) is the world's most comprehensive and broadly adopted cloud platform. We pioneered cloud computing and never stopped innovating — that's why customers from the most successful startups to Global 500 companies trust our robust suite of products and services to power their businesses. AWS Utility Computing (UC) provides product innovations — from foundational services such as Amazon's Simple Storage Service (S3) and Amazon Elastic Compute Cloud (EC2), to consistently released new product innovations that continue to set AWS's services and features apart in the industry. As a member of the UC organization, you'll support the development and management of Compute, Database, Storage, Internet of Things (IoT), Platform, and Productivity Apps services in AWS. Within AWS UC, Amazon Dedicated Cloud (ADC) roles engage with AWS customers who require specialized security solutions for their cloud services. Inclusive Team Culture: Here at AWS, it's in our nature to learn and be curious. Our employee-led affinity groups foster a culture of inclusion that empower us to be proud of our differences. Ongoing events and learning experiences, including our Conversations on Race and Ethnicity (CORE) and AmazeCon (diversity) conferences, inspire us to never stop embracing our uniqueness. Diverse Experiences: AWS values diverse experiences. Even if you do not meet all of the preferred qualifications and skills listed in the job description, we encourage candidates to apply. If your career is just starting, hasn't followed a traditional path, or includes alternative experiences, don't let it stop you from applying. Mentorship & Career Growth: We're continuously raising our performance bar as we strive to become Earth's Best Employer. That's why you'll find endless knowledge-sharing, mentorship and other career-advancing resources here to help you develop into a better-rounded professional. Work/Life Balance: We value work-life harmony. Achieving success at work should never come at the expense of sacrifices at home, which is why we strive for flexibility as part of our working culture. When we feel supported in the workplace and at home, there's nothing we can't achieve in the cloud.
US, VA, Arlington
PXT Central Science is seeking an exceptional Data Scientist to join our team. The ideal candidate will thrive in a dynamic, multifaceted role where you'll translate complex business challenges into rigorous quantitative frameworks, extract actionable insights from structured and unstructured datasets, and architect science-backed, scalable solutions that elevate the experience of our 1 million+ employees worldwide. If you're energized by the opportunity to apply data science to our mission of making Amazon Earth's Best Employer, we want to hear from you. Key job responsibilities • Own the design, development, and maintenance of scalable models and prototypes leveraging statistical, machine learning, or GenAI methodologies to enhance employee experience. • Partner with scientists, engineers, and product leaders to solve for employee experience defects using scientific approaches, building new services and tools that deliverable measurable impact. • Author and maintain detailed technical documentation related to the projects you drive. • Communicate results to diverse audiences of varying technical background with effective writing, visualizations, and presentations • Stay current with emerging methods and technologies, and implement them strategically to amplify the team’s impact. About the team The Central Science Team within Amazon’s People Experience and Technology org (PXTCS) uses economics, behavioral science, statistics, machine learning, and Generative AI to proactively identify mechanisms and process improvements which simultaneously improve Amazon and the lives, well-being, and the value of work to Amazonians. We are an interdisciplinary team, which combines the talents of science, engineering, and UX to develop and deliver solutions that measurably achieve this goal.
CA, BC, Vancouver
Building AI first solutions in a complex regulated space at Amazon scale. Would you like to work on one of the world's largest product classification and Tax computation engine? Or perhaps teaching AI to understanding nuances of domain specific language is what piques your interest? We are responsible for the tax calculation platform, providing the core services that calculate indirect taxes (sales tax and VAT) for all Amazon sales, physical and digital, globally. We are responsible for classifying billions of products in Amazon’s catalogue, calculating accurate Tax amount for every transaction that happens on Amazon’s global platform. Our engineering services are relied upon globally, built to be resilient across some of highest peak traffic demands in the world, while constantly evolving with the state of the art technology. We use language models at scale for Tax classification of the diverse products in Amazon catalogue and comprehending nuances of legislations and legal parlour.Our challenges include staying on top of the complex and fast-changing global tax legislations while ensuring computing calculations correctly and quickly, thousands of times a second, and each one needs to be accurate. Applied Scientists in Tax Engine works closely with Tax specialists, Product team , SDEs and Business teams - tackling a growing portfolio of core-science problems that includes balance of predictive and generative AI -including language comprehension, causal reasoning and active learning. As an Applied scientist, you will provide machine learning expertise to the team that helps increase the accuracy of Tax classification based on product information in Amazon catalogue. You will build various machine learning models to predict accuracy of tax classification, work with large reasoning models to reason with large volumes of system data and expert artefacts to teach domain intuition to AI. You will on variety of legislative changes acting as causal determinants to changes in our systems. You will help us innovate different ways to enhance for our global customers of Amazon.You will need to be entrepreneurial, work in a highly collaborative environment with SDEs, Product managers and businesses. We like to move fast, experiment, iterate and then scale quickly, thoughtfully balancing speed and quality. Whether you're passionate about taking your scientific solves to global reach and scale or one who seeks to solve complex problems under high scrutiny of an audited field, Tax Engine’s scope could offer you an intersting challenge to solve.
US, NY, New York
Prime Video is a first-stop entertainment destination offering customers a vast collection of premium programming in one app available across thousands of devices. Prime members can customize their viewing experience and find their favorite movies, series, documentaries, and live sports – including Amazon MGM Studios-produced series and movies; licensed fan favorites; and programming from Prime Video subscriptions such as Apple TV+, HBO Max, Peacock, Crunchyroll and MGM+. All customers, regardless of whether they have a Prime membership or not, can rent or buy titles via the Prime Video Store, and can enjoy even more content for free with ads. Are you interested in shaping the future of entertainment? Prime Video's technology teams are creating best-in-class digital video experience. As a Prime Video team member, you’ll have end-to-end ownership of the product, user experience, design, and technology required to deliver state-of-the-art experiences for our customers. You’ll get to work on projects that are fast-paced, challenging, and varied. You’ll also be able to experiment with new possibilities, take risks, and collaborate with remarkable people. We’ll look for you to bring your diverse perspectives, ideas, and skill-sets to make Prime Video even better for our customers. With global opportunities for talented technologists, you can decide where a career Prime Video Tech takes you! Key job responsibilities The science challenges include: - Heterogeneous customer responses. The same customer responds differently to ads depending on what they’re watching, how engaged they are, and their broader streaming context. We need to understand and predict this variation to make better advertising decisions in real time. - Signal sparsity in streaming. Unlike search or retail advertising, streaming offers no clicks, no conversions, and limited direct feedback. We must develop creative approaches to infer customer preferences, intent, and tolerance from indirect behavioral signals. - Personalization at scale. A one-size-fits-all ad experience leaves value on the table for both customers and the business. We build systems that adapt ad load, placement, and content to individual viewers across 100M+ customers and 100k+ titles. - Small effects, large variance. Ad interventions typically produce 0.1-2% shifts in engagement, effects easily overwhelmed by natural behavioral variance. Measuring, attributing, and optimizing these small signals requires rigorous experimental design and causal methodologies. - Competing objectives. Revenue, customer engagement, long-term retention, and advertiser value are in tension. We develop principled frameworks to navigate these tradeoffs and optimize for sustainable outcomes rather than any single metric. A day in the life - Lead the research and development of ML models that personalize advertising decisions for 100M+ customers across 100k+ titles, with production deployment in mind - Develop deep learning architectures (multi-task learning, embedding-based representations) for customer behavior prediction at scale - Design and analyze large-scale A/B experiments, applying causal inference techniques to measure and optimize the impact of ad strategies on customer engagement and monetization - Partner with engineering to ensure models meet production latency and scalability requirements - Collaborate with product managers to frame business problems as tractable ML problems and translate findings into product decisions - Shape the team’s scientific roadmap, identifying high-impact research directions About the team Prime Video Ads is pioneering the use of Generative AI, Deep Learning and Causal ML to empower the next generation of Ads experience in Prime Video. Our vision is to build the world’s most valuable and differentiated ad-supported streaming platform: one where every ad is relevant to the viewer, and every dollar spent by an advertiser delivers measurable results. We are seeking lead applied scientists who have demonstrated experience in advancing state of the art ML algorithms, particularly in Personalization, Relevance, Casual Inference and Gen-AI areas. Your role will be to deliver these innovations as production-ready systems at Amazon scale.
US, WA, Seattle
MULTIPLE POSITIONS AVAILABLE Employer: AMAZON.COM SERVICES LLC Offered Position: Applied Scientist II Job Location: Seattle, Washington Job Number: AMZ10564086 Position Responsibilities: Participate in the design, development, evaluation, deployment and updating of data-driven models and analytical solutions for machine learning (ML) and/or natural language (NL) applications. Develop and/or apply statistical modeling techniques (e.g. Bayesian models and deep neural networks), optimization methods, and other ML techniques to different applications in business and engineering. Routinely build and deploy ML models on available data. Research and implement novel ML and statistical approaches to add value to the business. Mentor junior engineers and scientists. Specialize in the field of Computational Advertising, directly advancing the techniques in Online Advertising Technology focusing on model optimization and latency reduction. Develop sophisticated machine learning algorithms and models using PyTorch. Develop advanced deep learning architectures to automate the discovery of complex, non-linear feature interactions, effectively capturing high-dimensional data patterns that traditional statistical models are unable to process. Investigate and implement high-order feature interaction architectures, such as DeepFM and DCN-V2, using PyTorch to capture complex non-linear relationships within sparse categorical datasets. Develop custom embedding layers in PyTorch to map high-dimensional, sparse features (e.g., Ad IDs) into dense, low-dimensional vector spaces. Conduct systematic research into the trade-offs between model architecture complexity and inference speed, utilizing PyTorch profiling tools to identify and optimize bottleneck layers. 40 hours / week, 8:00am-5:00pm, Salary Range: $163,617/year to $193,200/year Amazon is a total compensation company. Dependent on the position offered, equity, sign-on payments, and other forms of compensation may be provided as part of a total compensation package, in addition to a full range of medical, financial, and/or other benefits. For more information, visit: https://www.aboutamazon.com/workplace/employee-benefits. Amazon.com is an Equal Opportunity-Affirmative Action Employer – Minority / Female / Disability / Veteran / Gender Identity / Sexual Orientation.#0000
US, WA, Redmond
Amazon Leo is Amazon’s low Earth orbit satellite constellation initiative designed to bring fast, affordable broadband connectivity to unserved and underserved communities around the world. We are building a next-generation satellite system from the ground up — designing everything from the spacecraft and payload to the ground infrastructure and customer terminals. This is a rare opportunity to join a team that is shaping the future of global connectivity at scale. The Leo Antenna team is looking for an RF Systems Scientist who is excited to work at the intersection of radio frequency engineering, phased array antennas, large-scale data analysis, and hardware reliability. You will turn data from our customer terminal fleet into engineering decisions. Every terminal in the field reports a steady stream of telemetry: downlink and uplink signal quality, link state, antenna pointing, temperature, and reboots. Every terminal was also tested in the factory before it shipped, and the network sees every contact it makes with a satellite. Your job is to connect these sources. You'll find out which field behaviors trace back to the antenna hardware and which come from the installation, the sky around the terminal, the network, or a software release. You'll also look for patterns that point to weaknesses in the design. Your findings will shape how we design, build, test, and support the next generation of customer terminals. A typical problem starts small. A few terminals spend far more time recovering their link than their neighbors, run a little hotter than the rest of the fleet, or reboot more often than they should. You'll work out whether it's a real effect or noise, whether it's spreading, and what the affected terminals have in common. They might share a hardware revision, a software version, a location, or a signature in their factory test history. From there you'll form a hypothesis about the physical cause and argue it in front of antenna, RF, and software engineers who will push back hard. You'll also have to say plainly when the data rules your own idea out. To find the cause, you'll pull in the rest of each terminal's history: its factory and thermal test results, its software and reboot record, and what the satellite measured on the other side of the link. Some questions can be answered from data we already collect. For the rest, you'll decide what new measurement or test would answer them, work with the teams who can run it, and report what the results show and what should happen next. Beyond individual investigations, you'll build the fleet-level view. That means expected performance for a healthy terminal given where it is and what it sees, early warning signals for hardware that is starting to fail, and measures of how software releases and hardware revisions change the fleet. You'll decide which recurring analyses should become automated monitoring and work with software engineers to build them into our fleet health system. You'll be the person the team turns to when someone asks "is this terminal normal?" The ideal candidate understands how radio links and phased array antennas behave, is curious about how hardware fails, and is comfortable with messy data from many sources. They can turn "a few terminals look odd" into a rigorous analysis and a clear recommendation. Export Control Requirement: Due to applicable export control laws and regulations, candidates must be a U.S. citizen or national, U.S. permanent resident (i.e., current Green Card holder), or lawfully admitted into the U.S. as a refugee or granted asylum. Key job responsibilities Analyze customer terminal telemetry to measure antenna and link performance in the field, and detect failing hardware and slow degradation early Investigate underperforming terminals and field service requests, separating hardware effects from installation, obstruction, network, and software effects, and decide what should happen next Connect events across the fleet by hardware revision, manufacturing lot, software version, and location, and measure the effect of software releases and hardware changes on the fleet Correlate field behavior with factory and thermal test results, and feed what you learn back into test limits, screening, and the design of the next generation of terminals Define the fleet health metrics and baselines, specify the telemetry they need, and work with software engineers to turn recurring analyses into automated monitoring Build dashboards and AI-assisted tools that support fleet health reviews, field investigations, and design reviews Lead investigations that span several teams, from the first hypothesis to a decision, and present the evidence and recommendation to senior engineering leaders About the team The Leo Antenna team designs the phased array antenna inside Leo's customer terminals. It is the electronically steered panel that tracks satellites across the sky and carries every customer's connection. We own the antenna from simulation and design through factory test automation and thermal testing, and we own how it performs once it's installed at a customer's site. That includes the telemetry analysis and the fleet health system that keeps a daily hardware health record for every terminal in the field. We are a hardware-facing team that writes a lot of software, and we sit where several disciplines meet: antenna and RF design, terminal software, manufacturing, network operations, and field service. Every terminal has a part-level history, from its factory test results to every satellite contact it makes in the field. A large part of our work is investigation. When a terminal underperforms, we work out whether the cause is the hardware, the installation, the sky around it, the network, or the software, and we decide what should happen next. We have built analyses that map each terminal's view of the sky from its pointing and signal data, dashboards the wider organization uses every day, and AI-assisted tools that let engineers query telemetry and turn raw data into conclusions. We care about evidence over assumption. A ticket that lists a terminal doesn't explain it, and an obstruction request isn't an obstruction until the data shows one. Theories get tested against the data before they're accepted. Because we own both ends, from the first factory test to performance in the field, the loop is closed within the team. What we learn from the fleet goes straight back into our designs and test screens, so the next terminal we ship is more likely to work well for the customer who receives it.
US, WA, Seattle
Here at Amazon, we embrace our differences. We are committed to furthering our culture of diversity and inclusion of our teams within the organization. How do you get items to customers quickly, cost-effectively, and—most importantly—safely, in less than an hour? And how do you do it in a way that can scale? Our teams of hundreds of scientists, engineers, aerospace professionals, and futurists have been working hard to do just that! We are delivering to customers, and are excited for what’s to come. Check out more information about Prime Air on the About Amazon blog (https://www.aboutamazon.com/news/transportation/amazon-prime-air-delivery-drone-reveal-photos). If you are seeking an iterative environment where you can drive innovation, apply state-of-the-art technologies to solve real world delivery challenges, and provide benefits to customers, Prime Air is the place for you. Come work on the Amazon Prime Air Team! Prime Air's Flight Sciences High-Fidelity Methods (HFM) team is seeking an outstanding Applied Scientist to develop and verify drone systems models and flight physics models. These models form the backbone of every flight simulation performed within Prime Air, directly informing aircraft design, system verification, certification, and business decisions. The HFM team's work enables prediction of critical vehicle performance metrics—including range, maneuverability, tracking error, and aircraft stability—and connects vehicle design and operational decisions to business outcomes such as customer reachability. Our models are also a crucial input to the design of flight control algorithms and software verification. Because the accuracy and reliability of these models underpin so many facets of Prime Air's mission, the scientist in this role will have broad, high-visibility impact on the program's success. Key job responsibilities The Applied Scientist in this role will own the end-to-end lifecycle of simulation models—from development and deployment through verification and ongoing maintenance. This begins with gathering downstream customer needs, selecting the most suitable modeling approach and level of fidelity, and coordinating the generation of input data. It extends through model training, development and maintenance of software interfaces, and verification of model accuracy. A core aspect of this role is determining the right modeling approach for a given physical phenomenon. The scientist will need a working understanding of the physics and systems being modeled, including vehicle aerodynamics, propeller performance, multibody dynamics, atmospheric physics, electric powertrain components, and guidance and navigation system (GNS) sensors. They will design experiments to generate the data needed to build and verify models, and apply uncertainty quantification and statistical analyses to validate model fidelity.
US, CA, Palo Alto
Amazon Ads helps advertisers reach customers across Amazon's stores, Prime Video, Fire TV, Alexa, Twitch, and the open internet. Every ad we serve has to be safe, compliant, and trustworthy. The Ads Trust science team builds the AI systems that does this by reviewing billions of ads across text, image, video, audio, and product listings. We are looking for a Principal Applied Scientist to set the science direction for this work. This is an opportunity to apply as well as build state-of-the-art AI to problems where accuracy, scale, and customer trust all matter at once. This role suits a scientist who wants to own a hard problem end to end: inventing methods, proving them with strong evaluation, and putting them in front of customers. You will design the core modeling approach behind moderation at Amazon scale. That includes multimodal large language and vision-language models, retrieval-based labeling, model distillation for production, and customer facing agentic systems. You will take ambiguous, high-stakes problems, turn them into scientific roadmaps, and stay hands-on in the technical decisions that matter most. You will set the bar for scientific rigor, evaluation, and design reviews across a multi-team organization. You will also mentor senior scientists and partner with engineering, product, and policy leaders to get research into production. Key job responsibilities - Set the multi-year science strategy for automated content moderation, with high accuracy and efficiency as the goals. - Lead the design of multimodal and multilingual models for labeling and policy classification across text, image, video, audio, and product assets. - Drive methods that turn frontier-model quality into cost-efficient production systems, such as distillation, retrieval-augmented labeling, and efficient inference. - Define evaluation standards and measurement frameworks that set the quality bar for production models. - Own science low-level design reviews across teams and raise the technical bar for modeling, experimentation, and launch readiness. - Mentor and develop senior and mid-level scientists, and grow scientific leadership across the organization. - Influence engineering, product, and policy partners, and represent Amazon Ads science through publications and external collaborations.
US, CA, Irvine
Do you want to write the algorithms that control next-generation batteries, and shape the future of energy storage? This role sits where electrochemistry, data/ML, and control theory meet. You will develop the physics-informed and data-driven methods that estimate state of charge, health, power capability, and remaining life, and validate them against lab and fleet data until they are ready to ship. Your algorithms will decide how hard a battery can be pushed, when it needs attention, and how long it will last, so they shape the safety, warranty, and economics of the whole product. You will work side by side with cell, systems, and software engineers, and as an early hire you will help define how we develop and validate battery algorithms. This role can be based in Boston, Irvine, Sunnyvale, or Seattle, and involves occasional domestic and international travel, up to 10% of the time. Key job responsibilities • Develop and own algorithms for state of charge, state of health, power capability, and remaining useful life, combining physics-based models with data-driven and estimation methods. • Deliver validated reference implementations, test vectors, and accuracy targets that define what production code must achieve. • Design the test plans and data needed to develop and validate algorithms, working with the cell and testing teams to turn lab and field data into model improvements. • Analyze algorithm performance across the deployed fleet, find where models drift or fail, and drive improvements. • Own the accuracy and robustness of battery algorithms, and sign off that production implementations match the validated reference. Provide algorithm evidence for safety certification. A day in the life Your week moves between data, models, and experiments. One day you might be working out why a health estimate drifts on cells that have aged in the field, and the next deciding which model structure holds accuracy at cold temperatures. You could be designing a test with the cell team to expose a behavior your model misses, or reviewing an implementation with an applied scientist to confirm that what ships matches what you validated. You will regularly share results with software, systems, and cell leaders, and your findings will shape decisions well beyond the algorithm itself. About the team We are building Amazon's energy storage capability, providing power for advanced technologies while contributing to a stable grid. Our product is highly connected, yet secure and resilient by design. We own it end to end, from software, ML, and controls through system design to cell selection and supply, and on to a safety-certified product in the field. The organization operates like a startup inside Amazon: small, fast-moving, and accountable for our decisions. We work across hardware, software, and science, with suppliers and partners worldwide, and with Amazon infrastructure, procurement, legal, and finance. The organization is at an early stage and growing quickly. You will be among the first algorithm hires, so you will help shape how we design, build, certify, and ship, with a direct line of sight to the product and business decisions your work informs. This is a dynamic role that will regularly present new challenges as the product and organization evolve.
US, CA, San Francisco
The Amazon Center for Quantum Computing (CQC) is seeking an Applied Scientist to research and design cryogenic control hardware for large-scale quantum processors, with an emphasis on superconducting digital circuits. Controlling and reading out orders of magnitude more qubits than today's systems demands rethinking where classical control lives and what technology implements it; you will design and model superconducting digital logic (e.g., single-flux-quantum-family circuits) and related cryogenic electronics, evaluate them against the wiring, power, and thermal budgets of a scaled system, and co-design control approaches with the qubit devices they serve. You will carry designs from concept through simulation, layout, design review, and fabrication handoff, and your evaluations will inform which control-electronics directions the organization invests in. Key job responsibilities • Design and simulate superconducting digital circuits and cryogenic control/readout electronics for scalable qubit control • Develop and validate Josephson-junction circuit models, timing and margin analyses, and verification methodology for superconducting digital designs • Quantify system-level trade-offs — wiring count, power dissipation at each cryogenic stage, latency, footprint — across candidate control architectures, and communicate recommendations • Co-design control interfaces with qubit-device colleagues so control electronics and quantum devices are developed as one system • Deliver design artifacts through the organization's design-request and design-review processes, including tapeout-ready layouts and verification reports • Present results internally and, where appropriate, publish in peer-reviewed venues About the team The Enabling Technologies team develops technologies and relationships facilitating scaling, including alternative quantum processor architectures, superconducting qubit device packaging, and cryogenic control device electronics. We work alongside a co-located quantum processor design team focused on current-generation devices; the two teams share design and review processes, and moving ideas between them is an explicit part of our charter.