Data Engineer - Onboarding

Sardine is the leading agentic risk platform for fighting financial crime, and they are seeking a Senior Data/ML Engineer to own the data and machine learning foundation for compliance decisions. The role involves building and maintaining data pipelines, integrating new data sources, and productionizing ML models to enhance fraud detection and compliance processes.

Responsibilities:

• Own the data ingestion layer that brings device telemetry, transaction events, KYC/identity signals, and third-party enrichment into the platform — designing streaming pipelines (Pub/Sub, Apache Beam on Dataflow, Flink) and batch pipelines (Python, Airflow on Cloud Composer, Spark on Dataproc) that are correct, observable, and cheap to extend
• Build and evolve our feature platform, where the same Chronon feature definitions are computed by Flink for streaming and Spark for batch, with aggregation windows from one hour to 300 days, served to the rules engine and to models under a sub-second budget
• Establish feature correctness as an engineering discipline: streaming-versus-batch reconciliation, recomputation tests against the warehouse, train/serve parity checks, and drift monitoring that catches a broken feature before an analyst does
• Productionize fraud and identity ML models — training pipelines on Vertex AI and Kubeflow, gradient-boosted and tree-based models (XGBoost, LightGBM, CatBoost, scikit-learn), hyperparameter search, SHAP-based explanations, and score normalization — and build the automated retraining, champion/challenger promotion, and rollback machinery we don't yet have
• Engineer KYC, AML, and identity risk signals: document verification and doc-KYC outcomes, sanctions/PEP/adverse-media screening results, email and phone risk, synthetic identity indicators, bank and account verification, and periodic customer due diligence — turning noisy, multi-vendor, multi-jurisdiction data into features a model can actually learn from
• Integrate and harden new data sources, including 30+ third-party enrichment providers called in parallel on the request path, plus our cross-client consortium network — owning failover behavior, timeout budgets, graceful degradation, caching, and cost
• Own the warehouse and modeling layer in BigQuery — partitioning strategy, the staging-to-mart layer cake, training datasets, and the in-flight migration off dbt onto scheduled SQL and Python pipelines
• Design the entity resolution and graph data that link customers, devices, emails, phones, cards, bank accounts, and crypto addresses across clients, including large-scale connected-components work
• Make the platform safe by construction: field-level encryption for sensitive identifiers, regional data residency enforced in the pipeline definitions, PII handling and deletion paths, and feature-level gating so a bad signal can be turned off without a deploy
• Set technical direction and raise the team's ceiling — write the design docs, run the reviews, mentor engineers and data scientists, and decide what we build versus buy

Requirements:

• 8+ years building production data and ML systems, with real ownership of both the pipeline side and the model side. You have shipped models that made consequential automated decisions, not just dashboards
• Deep Python and strong SQL. You are fluent in a distributed processing framework (Spark, Beam, or Flink) and comfortable reasoning about streaming semantics — windowing, watermarks, late data, exactly-once versus at-least-once, and where correctness actually breaks
• Hands-on experience with a modern cloud data stack: GCP strongly preferred (BigQuery, Dataflow, Dataproc, Pub/Sub, Bigtable, Composer, Vertex AI) or the AWS equivalents, plus Docker, Kubernetes, Terraform, and CI/CD
• Practical ML engineering depth: feature stores and feature pipelines, training/serving skew, gradient-boosted tree models, class imbalance and rare-event modeling, threshold and cost-sensitive tuning, model monitoring and drift detection, and explainability
• Experience with high-volume, low-latency serving where a feature fetch has a few hundred milliseconds and there is no retry budget
• Domain experience in fraud, risk, payments, lending, or identity/KYC — or the demonstrated ability to get fluent in a regulated domain fast. You understand why label latency, feedback loops, and adversarial drift make fraud modeling different from ordinary supervised learning
• Comfort with data governance in a regulated environment: PII, encryption, access control, regional data residency, auditability
• Strong written communication. You can explain a modeling tradeoff to a fraud analyst and a pipeline design to a backend engineer, and you write things down
• A bias toward action and comfort in ambiguity. Much of this role is deciding what should exist, then building it
• Experience supporting customer-facing ML — bring-your-own-model integrations, model explainability for adverse action or regulatory review, or shadow/challenger scoring frameworks
• Experience in high-growth B2B SaaS, or as an early data/ML hire who built the function rather than inherited it