Key takeaways
- How Avantage evaluated around 40 machine learning tools and platforms before choosing H2O.ai, scored on cost, ease of productionizing models, and explainability
- How extending credit to previously rejected customers added an estimated 10,000 more advances a day in an early test, while keeping bad debt within target range
- How moving from a rules-based credit system to machine learning cut bad debt by roughly 19 percent
- Why explainability was decisive: the credit team reviews predictions example-by-example before every model launch to catch issues like data leakage
- What production machine learning demands beyond accuracy, including sub-0.3-second decisions at millions of transactions a day on systems that run on the telecom customers' own premises
- How automated feature drift monitoring let Avantage react quickly when customer behavior shifted during COVID
Overview
In this session, H2O.ai's John talks with Avantage's head of analytics and machine learning about moving airtime lending decisions from a rules-based credit system to machine learning in production. Avantage, a South African fintech that helps telecom companies issue financial products, evaluated around 40 tools before choosing H2O.ai. The conversation covers extending credit to an estimated 10,000 more customers a day in an early test, cutting bad debt by roughly 19 percent, and building explainability that a credit team could sign off on before every model launch.