Cutting mobile banking fraud by 35% for a pan-African operator using 100,000+ annotated transaction logs and multilingual LLM fine-tuning.

Client Overview

A major African mobile money operator needed to fortify its fraud detection system against rising digital scams. Existing models underperformed on transaction patterns unique to informal economies, prompting a need for fine-tuned LLMs to analyze user behavior and transaction narratives.

Project Scope

DataLens curated and annotated 100,000+ transaction logs, chat transcripts, and alert texts. Services included:

• •Named Entity Recognition (NER) for suspicious entities
• •Sentiment analysis on user queries
• •Sequence labeling for fraud pattern detection

LLM fine-tuning adapted open-source models for Swahili and Pidgin English contexts.

Challenges

• •High volume of multilingual, noisy data from SMS and app interactions.
• •Privacy compliance under evolving African data regulations.
• •Balancing false positives to avoid disrupting legitimate users.

DataLens Solution

Through end-to-end LLM fine-tuning, DataLens prepared domain-specific prompts and datasets, training models with techniques like LoRA for efficiency. Human annotators, trained in financial compliance, ensured GDPR-aligned de-identification while capturing socioeconomic nuances like informal remittances.

Results

• •Reduced fraud detection latency by 35% and false positives by 22%.
• •Fine-tuned model achieved 92% F1-score on validation sets.
• •Completed in 10 weeks, safeguarding $50M+ in annual transactions.