Can a large language model read a regulatory change and draft the requirements and control tests that follow from it, accurately enough for a Compliance Officer to sign off? How do you train a breach-risk model when confirmed breaches are rare, arrive months after the event, and mean different things to Compliance, Assurance, and the regulator? What does a model have to prove before a Compliance team acts on its output? We are looking for a Senior Applied Scientist to join the SPARC team within Risk and Compliance Solutions (RCS). SPARC's bold outcome is to accelerate the end-to-end regulatory change cycle by 10x, supercharging Compliance Officers as part of Amazon’s first AI-Native Second Line of Defense (2LoD). Our team builds the data, analytics, and modeling layer that RCS runs on, from regulatory reports and filings to the intelligence that tells Compliance Officers and our first line of defense (1LoD) partners where risk concentrates and which controls hold. You build the machine learning (ML) and generative artificial intelligence (GenAI) systems in that layer, take them into production, and own the evaluation science that lets Compliance rely on them. The work spans 3 areas: risk models that score which open risks and control weaknesses convert to breaches; GenAI systems that remove manual effort from requirement translation, control testing, and regulatory reporting, with human approval gates at every stage; and the evaluation science behind every model and agent the team ships. Financial crime compliance is the first domain, and its constraints shape the science: threats change faster than control sets adapt, and every output faces internal audit and external examination. Once an approach proves out, you carry it into the next compliance domain across the RCS portfolio. As a Senior Applied Scientist on SPARC, you own the science strategy for this layer, not just its individual models. You design and architect the agentic systems at the heart of it — objective-driven agents with human approval gates — and customize the technology to real compliance use cases. You define a multi-year science vision and roadmap for ML and GenAI in RCS, translate it into concrete plans for scientists and engineering partners, and anticipate where generative AI is heading so what you build capitalizes on compounding advances. You are the technical leader in AI/ML for the team and across partner teams, setting the standard for scientific rigor, raising the bar in design and model reviews, and mentoring scientists and engineers. You own problems from formulation to production: framing the science question with Compliance and 1LoD partners, designing experiments and backtests, building and validating models, and shipping them into the pipelines, reports, and tools that RCS teams use. The stack includes Amazon SageMaker, Amazon Bedrock, Amazon Redshift, and AWS Lambda, and your partners include data scientists, data engineers, business intelligence engineers, and software development engineers across RCS. The successful candidate treats an auditor's question as a design requirement, writes production-quality code alongside their models, and stays comfortable when the evaluation set does not exist until they build it. Key job responsibilities • Formulate ambiguous compliance problems as ML and GenAI problems with Compliance and 1LoD partners, and choose methods that fit the data, the risk, and the audit standard • Develop risk models that learn from rare, delayed, and noisy labels to score which risks and control weaknesses convert to breaches, and calibrate them so a score carries a defensible meaning • Build GenAI systems that remove manual effort from requirement translation, control testing, and regulatory reporting, with human approval gates at every stage • Design the evaluation for every ML and GenAI system you ship: golden sets that compliance experts label, metrics suited to rare outcomes and generated text, and release criteria a system meets before Compliance relies on it • Deploy and monitor models in production with software development engineers, including drift detection, performance tracking against realized outcomes, and retraining criteria • Explain model decisions, quantify uncertainty, and document methodology so Compliance Officers, auditors, and examiners can trace how a system reached each output • Create reusable science components, such as feature sets, scoring services, and evaluation harnesses, so an approach that works in financial crime compliance extends to the next domain without a rebuild • Design and architect the pod's agentic systems and customize them to compliance use cases, working toward 10x acceleration of the end-to-end regulatory change cycle while maintaining or improving quality • Set the technical direction for the pod's science: lead complex model and design reviews across teams, define evaluation methods and release standards others adopt, mentor data scientists and engineers on ML methods and experimental design, and publish or patent where the work and confidentiality allow • Define a long-term science vision for RCS compliance AI driven from Compliance and 1LoD needs, translate it into actionable plans for science, engineering, and product partners, and influence where the science impacts the portfolio next About the team Risk and Compliance Solutions (RCS) serves as a crucial safeguard for Amazon's buyers, brands, and selling partners while ensuring business teams meet their regulatory obligations. RCS operates as the second line of defense under the Three Lines of Defense (3LoD) model, providing specialized oversight, monitoring, and expertise while setting policies and procedures that guide Amazon's compliance framework. RCS's mission is to apply specialist industry insights and Amazon's unique approach to scale through technology, delighting customers through tailored compliance solutions. Payment regulatory compliance keeps the financial system closed to bad actors while widening access for hundreds of millions of good customers, and protects the licenses behind roughly $650B in annual activity across 18 entities and 45 regulated products in 6 regions.
August 18, 2018"Perfect hashing" is among the techniques that reduce the memory footprints of machine learning models by 94%.
