A Look Inside JPMorgan’s New Machine Learning Efforts

For the past few years, JPMorgan has been busy building out its machine learning capability under Daryush Laqab, its San Francisco-based head of A.I. platform product management, who was hired from Google in 2019. Last time we looked, the bank seemed to be paying salaries of $160,000-$170,000 to new joiners on Laqab’s team.

If that sounds appealing, you might want to watch the video below so that you know what you’re getting into. Recorded at the AWS re:Invent conference in December, it’s only just made it to you YouTube. The video is flagged as a day in the life of JPMorgan’s machine learning data scientists, but Laqab arguably does a better of job of highlighting some of the constraints data professionals at all banks have to work under. 

“There are some barriers to smooth data science at JPMorgan,” he explains.

A bank is not the same as a large technology firm. For example, data scientists at JPMorgan have to check that data is authorized for use, Laqab adds: “They need to go to a process to log that use and make sure that they have the adequate approvals for that intent in terms of use.”

They also have to deal with the legacy infrastructure issue: “We are a large organization, we have a lot of legacy infrastructure,” Laqab says. “Like any other legacy infrastructure, it is built over time, it is patched over time. These are tightly integrated, so moving part or all of that infrastructure to public cloud, replacing rule base engines with A.I./ML-based engines. All of that takes time and brings inertia to the innovation.”

JPMorgan’s size and complexity is another source of inertia, as multiple business lines in multiple regulated entities in different regulated environments all need to be considered. “Making sure that those regulatory obligations are taken care of, again, slows down data science at times,” Laqab advises.

And then there are more specific regulations, such as those concerning model governance. At JPMorgan, a machine learning model can’t go straight into a production environment. “It needs to go through a model review and a model governance process,” Laqab says. “To make sure we have another set of eyes that looks at how that model was created, how that model was developed.” And then there are software governance issues, too.

Despite all these hindrances, JPMorgan has already productionized A.I. models and built an ‘Omni AI ecosystem,’ which Laqab heads, to help employees to identify and ingest “minimum viable data” so that they can build models faster. Laqab says the bank saved $150 million in expenses in 2019 as a result. JPMorgan’s A.I. researchers are now working on everything from FAQ bots and chatbots to NLP search models for the bank’s own content, pattern recognition in equities markets and email processing. The breadth of work on offer is considerable. “We play in every market that is out there,” says Laqab.

The bank has also learned that the best way to structure its A.I. team is to split people into data scientists who train and create models and machine learning engineers who operationalize models. Before you apply, you might want to consider which you’d rather be.

A modified version of this article originally appeared in eFinancialCareers.