Staff Machine Learning Engineer - Risk Intelligence / AI Agents

About the role

OKX is hiring a Staff Machine Learning Engineer for its risk intelligence function in San Jose, California. The role builds machine learning systems for risk at a global crypto exchange, covering on-chain activity, fiat deposits and withdrawals, trading behaviour, device intelligence, identity and customer interactions. It sits in a multidisciplinary risk team and works across the full ML lifecycle, from problem formulation and feature engineering to real-time deployment, monitoring and experimentation. The engineer also helps shape how AI is used across the risk organization, including LLM-assisted development, automated model workflows, AI-powered investigations and intelligent review agents.

Key responsibilities

• Design, build and deploy ML models for payment fraud, account takeover, scam detection, deposit and withdrawal risk, promotional abuse, customer risk assessment and transaction monitoring.
• Own production ML systems end to end, including feature pipelines, training workflows, model serving, decision integrations, monitoring, alerting, drift detection, retraining and incident response.
• Partner with risk strategy and product teams to turn models into production controls such as approvals, rejections, reviews, cooldowns, limits and account restrictions.
• Work with risk operations to understand investigation workflows, incorporate reviewer feedback, improve model explainability and refine labels and training data.
• Use LLM coding tools throughout the workflow while keeping security and review standards.
• Develop AI-powered capabilities such as investigation agents, case summarization, evidence collection, alert triage and automated decision support.
• Bring research-stage models into production by validating feature logic, checking data quality and ensuring consistency between offline training and online inference.
• Make models explainable, traceable and well documented for risk operations, product stakeholders, governance teams and regulators where applicable.
• Design and deploy LLM-based agents for case triage, evidence retrieval, transaction analysis, alert summarization and automated action orchestration, using tool calling, RAG, workflow orchestration, guardrails and human-in-the-loop controls.
• Build evaluation frameworks for LLM agents (accuracy, task completion, latency, cost, reviewer acceptance) and ensure safety through permissions, audit logs, data privacy, prompt and tool security, fallback mechanisms and escalation paths.

Required profile

• Significant professional experience in machine learning engineering, applied data science or a closely related field, with a strong record of moving models from prototype to production; scope and level are calibrated to experience.
• Demonstrable fluency with AI-assisted engineering: regular use of LLM coding tools, built AI-integrated workflows or applications, and awareness of their security, reliability and governance risks.
• Experience integrating LLM agents with internal systems, APIs, databases, search tools, case-management platforms or decision engines.
• Experience building AI agents for fraud, risk, compliance, customer operations, cybersecurity or similar high-stakes domains is an advantage.
• Strong communication and collaboration skills with engineers, data scientists, risk specialists, product managers, operations teams and legal or compliance stakeholders.

Required skills

• Strong Python and hands-on experience with PyTorch, TensorFlow, XGBoost, LightGBM or scikit-learn.
• Supervised learning, anomaly detection, representation learning, class-imbalanced modeling, model calibration and evaluation under changing data distributions.
• Familiarity with SHAP, feature attribution, reason-code generation and model scorecards.
• Hands-on experience with production LLM agents, including agentic workflows, tool calling, RAG, prompt and context management, structured output and multi-step task orchestration.
• Understanding of LLM-agent evaluation and reliability, including hallucination control, grounding, observability, permissions, failure handling, human review, latency and cost optimization.