[Remote] Staff Data Scientist

Note: The job is a remote job and is open to candidates in USA. Sift is an AI-powered fraud platform securing digital trust for leading global businesses. They are seeking a Staff Data Scientist to architect advanced modeling strategies across fraud and abuse problem domains, leveraging deep expertise in fraud patterns and statistical rigor to enhance model performance and prevent fraud.

Responsibilities

• Architect and own advanced modeling strategies across fraud and abuse problem domains (payment fraud, account takeover, identity spoofing, account abuse, content manipulation, credential stuffing). Your deep understanding of attacker tactics, exploit chains, and evasion strategies informs which signals matter and which are noise. You'll drive framework selection—deciding when gradient boosting on velocity features suffices, when graph neural networks unlock network effects competitors miss, when deep learning on sequence data catches adaptive fraud patterns—and hold yourself accountable for production outcomes. You'll work backward from business metrics (customer adoption, chargeback reduction, operational lift) to model objectives •informed by threat models•
• Establish and defend model quality standards that account for adversarial dynamics. You'll develop diagnostic frameworks to decompose model performance by fraud type, attacker sophistication level, geography, and temporal patterns. You'll own the post-launch monitoring process, identify when degradation signals retrain vs. architecture change vs. active evasion by fraud rings. You'll design sampling strategies that catch emerging fraud patterns before they scale. Your infosec intuition becomes your quality moat: you'll spot when performance drops aren't random—they're a signal that attackers have found a new exploit path
• Lead statistical innovation on our highest-leverage fraud problems. You'll explore novel feature representations drawn from your understanding of fraud mechanics (network propagation of compromised accounts, timing signatures of automated attacks, behavioral deviation from account history). You'll run rigorous experiments to validate whether a suspected fraud pattern is exploitable or a false lead. You'll publish findings internally (and externally where disclosable), and mentor junior data scientists on the difference between statistical significance and security-relevant signal magnitude
• Partner with ML engineering and information security on adversarial robustness. You'll co-design models that don't just maximize accuracy—they resist manipulation. You'll pressure-test feature importance against known evasion tactics. You'll own the handoff from research to serving, ensuring what ships hasn't leaked assumptions about attacker behavior. Your infosec depth means you're fluent in threat modeling conversations with security teams, not learning it on the job
• Build automated workflows that scale human expertise while respecting fraud complexity. You'll leverage AI-assisted tools (LLMs, AutoML frameworks) to accelerate experimentation while maintaining verification checkpoints informed by your domain knowledge. You'll document which automation patterns you trust (feature engineering exploration) and which require human oversight (fraud strategy pivots that might break assumptions in your features). You'll become the SME on where humans and AI each belong in fraud modeling pipelines

Skills

• Deep, hands-on knowledge of fraud and information security patterns. You've modeled payment fraud, account takeover, identity abuse, or network attacks in production. You understand attacker incentives, exploit chains, evasion tactics, and how fraud patterns evolve in response to defenses. You can explain the difference between a velocity signal that's correlated with fraud and one that's causal—and why attackers can't simply game it. You're not learning fraud from blog posts; you're bringing operational context from having debugged production systems under attack
• 5+ years of hands-on modeling experience with production accountability. You've shipped models to millions of users, owned their performance in production, and made decisions based on what's broken and why—not just benchmark scores. Bonus: some of that experience comes from adversarial or security-adjacent domains
• Deep expertise in multiple modeling paradigms: Tree-based methods (XGBoost, LightGBM with parameter mastery), deep learning architectures (CNNs, RNNs, transformers for sequential/graph data), and graph-based methods (GNNs, message passing, network propagation). You know when each is overfit versus underspecified. You've chosen frameworks based on problem structure, not trend
• Advanced degree in Statistics, Data Science, Machine Learning, or equivalent (MS or PhD in quantitative field, or 8+ years of demonstrable statistical modeling depth in production fraud/security contexts). You should reason naturally about confidence intervals, bias-variance tradeoffs, and statistical significance—not just memorize formulas. We care more about statistical intuition + proven execution than pedigree
• Lean, deep statistical intuition informed by domain reality. You can explain why a fraud model is failing through first principles (feature leakage from attacker behavior that changed, distribution shift from geography expansion, optimization pathology from class imbalance). You spot when a metric is gaming the objective. You know the difference between a model that's broken and one that's working correctly but facing a new fraud strategy
• Proven ability to partner with AI-assisted automation tools. You use LLMs, AutoML, and other AI systems to move faster—especially for feature engineering exploration and pattern discovery—but you verify their outputs and know where they hallucinate or oversimplify. You're building intellectual scaffolding, not outsourcing judgment. Fraud modeling can't be delegated to automation; you're the gate
• Comfort working in ambiguity and adversarial contexts. You don't wait for perfect specs—you clarify what 'reducing fraud leakage' means for a specific customer, run a small experiment, present findings with uncertainty bands, and iterate. You're comfortable saying 'attackers might exploit this assumption' or 'we need more data on this vector.' You're comfortable saying 'this is a business decision about fraud tolerance, not a modeling decision.'

Benefits

• Offers Equity

Company Overview

Company H1B Sponsorship