Staff+ Software Engineer (Account Abuse, Machine Learning)

About this role

• The Account Abuse team is tasked with ensuring Anthropic’s computing capacity is allocated fairly, minimizing resources available to bad actors and preventing them from coming back
• As a software engineer on this team, you will build the machine learning systems that help us detect and stop abuse at scale
• The ideal candidate can see things from opponents’ perspectives, understand their means and motives, and anticipate their responses to countermeasures
• We’re looking for full stack machine learning engineers with experience across model training, productionization, and evaluation
• You’ll also look for ways to use Claude to speed up how these models get built and maintained
• This is classical ML on structured and behavioral data
• You do not need a deep learning background or knowledge of LLM internals
• What matters is that you have trained and shipped models where the stakes are real, and that you care about building robust production systems as much as the model itself
• A false positive here is a legitimate customer locked out, so measurement, precision, and safe rollout are part of the job
• Build and operate a feature computation platform that serves both model training and real-time scoring, with point-in-time correct training data and low-latency online retrieval
• Train, evaluate, and deploy models that detect account-level abuse and fraud, running them both offline and online
• Build tooling that automates more of the model development lifecycle, including using Claude to speed up feature development, training, and evaluation
• Make backtesting, shadow deployment, and staged rollout the default path to production, with monitoring for training / serving skew, drift, and adversarial adaptation
• Work with our data scientists and our Policy & Enforcement team to improve label coverage and quality
• Partner with product and platform teams to gather signals and integrate model decisions with minimal impact on their systems’ latency, stability, or overal...