Our CTO Tells Us the Surprising Things He’s Learned Along the Way
Latest posts by Paul Bookstaber (see all)
- A New Timescale for Fraud Science: Insights from Our CSO - October 17, 2017
- Operationalizing Machine Learning for Fraud: Human + Machine Learning is the Best of Both - October 6, 2017
- Easing Into Machine Learning for PSD2: An Interview with Our VP of French Sales - September 13, 2017
Over the past few years, from his vantage point as Feedzai’s head technologist, CTO Paulo Marques has seen the question change. For leaders at major banks, the original question was: “Do I need machine learning?” Now that the answer is yes, the new question is: “How can I operationalize machine learning?”
Accenture surveyed 5,600 companies and found that 79% of them are making either “moderate or “extensive” investments in AI capabilities over the next three years. Where these companies diverge is in the strategy for whether to build or outsource the AI technologies necessary to develop these capabilities. 14% of the companies surveyed are developing AI in-house, using their own resources. 13% of companies are going to external vendors, like AI fintech startups. 17% are doing both, combining in-house resources with external collaboration.
As companies begin to invest heavily in AI, including building their own teams of data scientists and AI engineers, they are grappling with sets of decisions about which technologies to build, and which technologies to buy. It’s a debate that Mr. Marques has seen unfold several times in his discussions with enterprises seeking a machine learning platform for risk.
One nexus for debate is in the relationship between open source technology and proprietary technology. For example, Feedzai use 132 open source tools and libraries in its proprietary system. It’s a fact that leads some executives to wonder why they can’t leverage the same open source technologies with in-house teams to achieve similar ends. These executives will explore the option of leveraging in-house AI expertise to build a risk platform themselves. “I haven’t seen an organization succeed yet,” Paulo writes.
To demonstrate some of the unseen obstacles on the road toward building a machine learning platform for risk, Paulo published the report below, which you can download here.
The specifications for a high-performance risk platform can be prohibitively demanding. Here’s one example, in Paulo’s words:
“The car we’re building at Feedzai needs to fulfill certain criteria in terms of efficiency, power and accuracy. In Portugal we have minimum speed rules on our highways. You need to go above 50 kilometers an hour. Imagine you’ve spent years building a platform, only to learn that it didn’t meet latency requirements, that your car didn’t drive at the speed minimum. That’s what happened with one large American bank I know: they bought a platform, spent over a year to build the input integration for scoring, and by the time they put it into production, they learned their latency was too high.”
Read the full report for an explanation of the five misconceptions:
- “I have the parts, so I can build the whole”
- “Our industry has a universal definition for ‘complete'”
- “I can build this in a year”
- “What works in the sandbox will work in the field”
- “Metrics are the end, not the beginning”