Acquiring fraud detection. Less friction. Less loss. Governed AI.

What I can deliver

• Time aware evaluation protocol, rolling windows, leakage controls, reproducible reporting.
• Cost model and threshold policy: net benefit curves, friction budget, capacity constraints.
• Model benchmarking: strong baselines first, deep learning only when it wins under temporal backtests.
• Governance pack: model cards, release gates, monitoring plan, rollback triggers, audit evidence.

Practical scope. Acquiring fraud, card present and e-commerce. Design for regulated environments.

Engagement modes

• Advisory review of your current stack and evaluation practice. Findings, risks, concrete fixes.
• Hands on build of the evaluation harness and governance templates, ready for production use.
• Benchmark sprint: baseline suite, temporal CV, cost curves, recommendation with evidence.
• Enablement: train analysts and engineers to run the pipeline and keep it honest under drift.

You keep ownership. I deliver the method, artefacts, and evidence trail.

Credibility signals

This site is the public face of the method. Private engagements remain private. Nothing sensitive is published here.

Research questions

• How can we reduce false positives while preserving or improving fraud capture under strict time aware validation?
• How should we evaluate models when the real target is net benefit, not accuracy?
• How can we detect and control drift in acquiring, where merchant and attacker behaviour evolves continuously?
• How can governance, auditability, and human oversight be operationalised without reducing performance?

Scope. Card present and e-commerce acquiring. Merchant context. Behavioural signals. Cost aware decision policies.

Facts and figures

Core outputs on this hub

Evaluation protocol

Fraud data is non stationary. Random splits are misleading. This work uses blocked time series validation and rolling windows to approximate production reality, quantify drift sensitivity, and prevent leakage.

PDF. Download

Cost model note

The primary KPI is expected net benefit. False positives create friction and operational cost. False negatives create direct loss. This note defines assumptions, constraints, and reporting requirements for threshold selection.

PDF. Download

Model card template

Audit ready documentation. Model overview, data windows, temporal results, drift monitoring, governance sign off, and rollback criteria.

PDF. Download

Governance principles

• Data minimisation and explicit purpose limitation for anti fraud.
• Documented lineage. Data version, features, training window, parameters.
• Human oversight for high impact decisions with clear accountability.
• Auditability by design. Logs, model cards, decision traceability.

Operational explainability

Explainability is treated as an operational instrument. The goal is to enable analysts, risk teams, and auditors to understand why a decision occurred and how stable that reason is under drift.

Policy. Privacy and data disclosure

Disclosure statement

No production secrets, sensitive rules, merchant identities, or personal data are published. Public materials use synthetic data or anonymised aggregates. Any public robustness demonstrations are designed to inform defence, not enable abuse.