Senior Data Scientist

Empirical Security; Chicago (IL); Sur place; USD 120 000 - 180 000; Plein temps

Résumé du poste

Empirical Security is seeking a Security Data Scientist to own end-to-end modeling for vulnerability predictions. You will design, train, and deploy models using ground-truth telemetry across cloud, appsec, and traditional environments.

Expect to tackle extreme class imbalance and to explain results to executives with clear uncertainty quantification. The role involves building features from scanner output, EDR, asset inventory, identity, and cloud posture while ensuring robust monitoring and

Qualifications

• Proven ML models in production with real impact.
• Fluency in Python and SQL with version control discipline.
• Experience handling extreme class imbalance and advanced methods (survival/Bayesian/causal).
• Ability to explain uncertainty and model results to executives.

Responsabilités

• Design, train, and deploy exploit-prediction models.
• Evaluate models for precision, recall, coverage, calibration.
• Address extreme class imbalance in CVE exploitation data.
• Develop hierarchical, partial-pooling models for cold-starts.
• Engineer features from scanner output, EDR, assets, identity, cloud posture.
• Monitor drift and maintain models in production.
• Publish methodology, benchmarks, and open results.
• Collaborate with engineering to move models into production systems.

Connaissances

Applied ML

Python

SQL

Version control

Reproducible pipelines

Explainability

Uncertainty quantification

Model deployment

Outils

Git

Jupyter

Python libraries

Description du poste

Empirical Security is seeking an experienced Security Data Scientist focused on building the next generation of cybersecurity vulnerability models. Our unique approach leverages ground-truth telemetry to develop predictive, actionable insights that transform the way organizations identify, prioritize, and remediate vulnerabilities in cloud, appsec and traditional environments. We build models specific to individual customers, and maintain many of them side by side.

The role

You own models and data end to end: problem framing, features, training, evaluation, deployment, and the uncomfortable part where you explain to a customer why the vulnerability their board is worried about ranked 400th on the remediation list.

What you'll do

• Design, train, and ship exploit prediction models against ground-truth exploitation telemetry, in cloud, appsec, and traditional infrastructure.

• Build evaluation that survives contact with reality. Precision, recall, coverage, efficiency, calibration, and how all four decay over time. Accuracy is not a number you report once at launch.

• Solve for extreme class imbalance. A fraction of a percent of published CVEs are ever exploited in the wild, and most of the industry's modeling failures start with pretending that isn't true.

• Work the hard part of the dual-model architecture: partial pooling, hierarchical priors, and cold-start behavior for customers whose local telemetry is thin in month one and rich in month twelve.

• Engineer features across scanner output, EDR, asset inventory, identity, cloud posture, and exploitation telemetry, and be honest about which ones are leakage.

• Own monitoring and drift detection.

• Publish. Papers, methodology write-ups, open benchmarks, conference talks. Our positioning is that we show our work.

• Partner with engineering and our forward deployed team to move models out of notebooks and into production systems that customers depend on.

What you'll need

• Several years of applied machine learning or statistics with models that ran in production and had consequences when they were wrong.

• Fluency in Python and SQL, and the discipline that comes with version control, reproducible pipelines, and secure handling of customer data.

• Real depth in classification under heavy imbalance, plus at least one of: survival and time-to-event analysis, Bayesian hierarchical modeling, or causal inference.

• Calibration instincts. You should be visibly uncomfortable when a model outputs 0.9 and is right 60% of the time.

• The ability to explain a model to a security executive, and to quantify uncertainty out loud rather than burying it in an appendix.

• Enough curiosity about attacker behavior to ask why a feature works, not just whether it does.