AI Development Company in Miami

Fintech / Payments

A Miami-based fintech company ($890 million in annual payment volume, serving 45,000 small businesses that send and receive cross-border payments between the US and Latin America — primarily Mexico, Colombia, Brazil, and Argentina) was losing $8.4 million annually to fraud. The company's fraud detection was a rule-based system supplemented by manual review. Problems: the rule-based system generated 12,000 alerts daily but only 340 were genuine fraud (2.8% precision — 97.2% false positives), draining the 15-person compliance team. The system missed sophisticated fraud patterns: invoice manipulation (fraudulent invoices with slightly modified bank details — the real supplier's invoice intercepted and the bank account changed), account takeover (compromised business accounts used to initiate fraudulent transfers — difficult to detect because the transactions appeared to come from legitimate businesses), and velocity abuse (fraudsters making many small transactions below the review threshold — individually normal, collectively fraudulent). Cross-border complexity made fraud detection harder: legitimate transaction patterns vary dramatically by country (a $50,000 payment to a Colombian manufacturer looks very different from a $50,000 payment to a Mexican distributor), currency conversion added noise (exchange rate fluctuations made amount-based rules unreliable), and LATAM banking infrastructure created latency (some LATAM bank confirmations take 24-48 hours, creating a window for fraud). The company also faced regulatory pressure: FinCEN had issued guidance specifically about cross-border payment fraud, and the company's BSA examiner flagged the high false positive rate as evidence of an "ineffective" monitoring programme.

We built a multi-model fraud detection system designed for cross-border payment patterns between the US and Latin America. (1) Transaction-level model: a gradient-boosted ensemble trained on 2 years of transaction data (4.8 million transactions, including 3,200 confirmed fraud cases). Features specific to cross-border: corridor-specific patterns (US→Mexico transfers have different normal patterns than US→Brazil — the model learns corridor-specific baselines), beneficiary analysis (comparing the receiving bank account, beneficiary name, and country against the business's historical payment patterns — detecting account detail changes that indicate invoice manipulation), velocity by corridor (normal transaction frequency varies by country — a business that sends 20 payments/month to Mexico but suddenly sends 5 to Argentina in a day triggers investigation), FX-adjusted amounts (normalising transaction amounts by exchange rate to detect amount manipulation — a fraudster increasing an invoice by 15% might be masked by a 10% currency movement), and business profile deviation (each business has a normal pattern — countries served, average amount, payment frequency — deviations from this profile scored by magnitude and risk). (2) Network analysis: a graph model connecting businesses through shared beneficiaries, bank accounts, IP addresses, and device fingerprints. The graph reveals: money mule networks (accounts that receive funds from multiple compromised businesses and quickly forward to other accounts — the graph pattern is distinctive), shared infrastructure (multiple businesses using the same IP address or device to initiate payments — potentially indicating a single fraudster operating multiple accounts), and beneficiary clusters (groups of beneficiary accounts at the same bank, opened around the same time, receiving payments from different businesses — a common pattern in organised fraud). (3) Invoice manipulation detection: a specialised model for the most common fraud type. Invoice comparison: when a payment instruction references an invoice, the system compares the beneficiary details against the business's historical invoices to the same supplier — flagging changes in bank account, routing number, or beneficiary name. NLP for invoice analysis: extracting and comparing key fields from uploaded invoices across payments — detecting subtle modifications (a single digit changed in a bank account number, a slightly different beneficiary name). Email compromise indicators: when invoice manipulation is suspected, the system analyses the email trail (with customer consent) for signs of business email compromise — unusual sender addresses, forwarded invoices from external accounts, or changes in email communication patterns. (4) Adaptive thresholds by corridor: rather than one-size-fits-all rules, the system maintains dynamic thresholds per corridor. The US→Mexico corridor has different normal patterns than US→Argentina. Thresholds adjust seasonally (agricultural payments to Mexico peak in harvest seasons), by business type (importers have different patterns than service companies), and by current risk environment (when fraud attempts increase in a specific corridor, thresholds tighten automatically). (5) Compliance integration: every fraud alert generates a structured report formatted for BSA/AML compliance. Suspicious Activity Reports (SARs) are auto-drafted from confirmed fraud cases — including narrative, transaction details, and supporting evidence — reducing SAR filing time from 4 hours to 45 minutes per report.

Fraud detection: precision improved from 2.8% to 34% (12x improvement) — alerts decreased from 12,000/day to 1,800/day while catching 23% more fraud. The compliance team now reviews 1,800 meaningful alerts instead of 12,000 noisy ones. Fraud losses: decreased from $8.4M to $3.1M annually (63% reduction, $5.3M saved). Invoice manipulation losses decreased 78% (the specialised invoice detection model caught the highest-value fraud type). Network analysis identified 3 organised fraud rings totalling $1.2M in prevented losses — patterns invisible at the individual transaction level. False positive reduction: from 97.2% to 66% — still improving but dramatically more actionable. The compliance team's effective capacity increased 6x (reviewing 1,800 alerts at 34% precision vs. 12,000 at 2.8% produces 612 genuine investigations vs. 336 — more fraud caught with less work). SAR filing: auto-drafted SARs reduced filing time from 4 hours to 45 minutes. The compliance team filed 40% more SARs (catching more reportable activity while spending less time per report). BSA examiner: the improved precision and SAR quality were positively noted in the next BSA examination — the examiner upgraded the monitoring programme rating from "needs improvement" to "satisfactory." Corridor-specific accuracy: US→Mexico (highest volume): 89% recall at 38% precision. US→Colombia: 85% recall at 31% precision. US→Brazil: 82% recall at 29% precision. The model performs best on the highest-volume corridors where training data is richest. Regulatory value: beyond fraud prevention, the improved compliance programme protects the company's money transmitter licences in 48 states — licence revocation would be an existential business risk. Development investment: $360,000 over 16 weeks.