AI Development in Sydney, Australia

Financial Services / Banking

A Sydney-based digital bank (ADI-licensed, 280,000 active customers, A$3.8B in deposits, processing 4.2M transactions monthly) needed to upgrade its fraud detection from rules-based to AI-powered. The problems: (1) Rules-based limitations: the existing system used 340 fraud rules (if transaction > $X from country Y, flag for review). These rules caught known fraud patterns but missed novel attacks. False positive rate: 4.2% — meaning 176,000 legitimate transactions monthly were incorrectly flagged, each requiring manual review. The fraud operations team (8 analysts) spent 85% of their time reviewing false positives rather than investigating actual fraud. (2) Real-time requirement: fraud detection had to evaluate each transaction in under 200ms — the time between transaction authorisation request and response. The rules engine achieved this, but adding more rules increased latency. At 340 rules, the engine was already at 180ms — approaching the limit. (3) Emerging fraud patterns: account takeover attacks (phishing → credential compromise → fraudulent transactions) had increased 300% year-over-year. The rules engine couldn't detect ATO because the transactions originated from the legitimate customer's device profile — only the behavioural pattern (unusual transaction timing, amounts, or recipients) indicated fraud. (4) APRA expectations: APRA's thematic review of fraud management highlighted that ADIs should demonstrate: model governance (validation, monitoring, documentation), explainability (ability to explain why a transaction was flagged), and ongoing performance measurement. The bank needed an AI approach that satisfied these regulatory expectations.

We built a real-time AI fraud detection system over 14 weeks: (1) Feature engineering: 280+ features computed in real-time for each transaction — transaction-level (amount, merchant category, country, time of day), customer-level (average transaction amount, typical merchants, usual transaction times, historical fraud rate), device-level (device fingerprint, IP geolocation, session behaviour), velocity features (transaction count in last 1hr/6hr/24hr/7d, unique merchants in last 24hr, unique countries in last 7d), and network features (is the recipient a known mule account, has this merchant been associated with fraud, does this payment network have elevated fraud rates). Feature computation: sub-10ms using Redis for pre-computed customer profiles and real-time aggregation. (2) Ensemble model: three models running in parallel — a gradient-boosted tree (XGBoost) for transaction-level anomaly scoring, a sequence model (LSTM) for behavioural pattern analysis (detecting ATO by identifying unusual sequences of actions), and a graph neural network for detecting money mule networks (analysing transaction relationships to identify accounts used for laundering). The ensemble: combining scores from all three models using a calibrated meta-model — producing a single fraud probability per transaction. (3) Real-time inference: model serving on AWS ap-southeast-2 (Sydney) via FastAPI with Redis caching. Inference latency: 35ms average (well within the 200ms budget — leaving room for network transit and downstream processing). Throughput: handling 4.2M monthly transactions with horizontal scaling during peak periods (Friday afternoons, pay day, Black Friday). (4) Explainability: SHAP (SHapley Additive exPlanations) values computed for every flagged transaction — identifying which features contributed most to the fraud score. Fraud analysts receiving: the fraud probability, the top 5 contributing factors (e.g., "unusual country for this customer," "velocity exceeds 99th percentile," "device fingerprint not previously seen"), and a recommendation (block, review, allow with monitoring). APRA documentation: model governance framework including model validation report, ongoing monitoring dashboard, and model risk assessment. (5) Feedback loop: analyst decisions (confirmed fraud, false positive) feeding back into the model training pipeline — the model improving continuously as it learns from analyst expertise.

False positive rate: from 4.2% to 0.8% (81% reduction). Monthly false positives: from 176,000 to 33,600 — each false positive requiring 5 minutes of analyst time, this saved 11,900 hours annually. Fraud detection rate: from 72% (rules engine) to 94% (AI model) — catching an additional 22% of fraud that the rules engine missed. Dollar impact: A$2.8M in prevented fraud in the first 12 months (fraud that the AI caught but the rules engine would have missed — measured by retroactive analysis of transactions the rules engine would have scored below threshold). Account takeover detection: the LSTM behavioural model detecting 89% of ATO attacks — a fraud category the rules engine had zero capability for. ATO losses: reduced from A$1.2M annually to A$180,000 (85% reduction). Response time: inference latency at 35ms (vs. 180ms for the rules engine) — actually faster despite the far greater analytical depth. This headroom: allowing the bank to add new features without approaching the 200ms limit. Analyst efficiency: the fraud operations team transitioning from 85% false-positive review to 60% investigation of genuine fraud attempts. Analyst job satisfaction: improved significantly (investigating actual fraud is more engaging than reviewing false alarms). One analyst: identified a previously unknown fraud ring through investigation time freed by reduced false positives — this single investigation preventing A$400,000 in potential losses. APRA: model governance documentation reviewed during supervisory examination — no adverse findings. The SHAP explainability implementation: specifically noted as meeting APRA's expectations for model transparency. Development cost: A$340,000 (14 weeks). Annual value: A$2.8M (additional fraud prevented) + A$1.0M (ATO reduction) + A$600,000 (analyst efficiency — 11,900 hours × A$50/hr equivalent) = A$4.4M annual value. ROI: development cost recovered within first 5 weeks from fraud prevention alone.