In an interview last year, Dhir said vaguely that they focus on, "cutting edge conversational AI (large AM/LM) systems to work in the finance & banking commerce domain.”

JPMorgan's West Coast technologists are among its best paid. They're also famously drawn from the likes of Facebook, Google and Apple. But the moves aren't all one way - JPMorgan has lost staff to the tech giants while simultaneously adding a raft of talent to its new team developing chatbots in California. 

Last November, JPMorgan's San Francisco office hired Chandra Shekhar Dhir, a machine learning manager who previously worked on Apple's 'Hey Siri' chat device for five and a half years. This month, Manasa Hari, a head of product management and CTO of JPM's data platform in California, moved in the opposite direction. Hari quit JPMorgan and went to Apple in Cupertino as head of product management for the AI/ML training platform.  

Some of JPMorgan's highest-paid technologists are in California. After all, this is where the bank's VP-level developers are on salaries alone of $317k, and where JPMorgan's associate-level developers have been spied on salaries of $260k. Those are strong salaries compared to pay at Apple, where H1B salary data says machine learning engineers are on $230k salaries and 'acoustic machine learning specialists' have been hired on 'just' $160k. 

JPM's high California salaries go some way to explaining Dhir's success in building out his new team in San Francisco. Since he arrived at JPMorgan in November, Dhir's hired Jaekwon Yoo, a senior machine learning engineer from PlayStation, Anh Nguyen, a junior engineer from Microsoft and Peter Plantinga, a former researcher at the Quebec Artificial Intelligence Institute (who's based in Ohio). 

The chatbot team are all about audio AI. In an interview last year, Dhir said vaguely that they focus on, "cutting edge conversational AI (large AM/LM) systems to work in the finance & banking commerce domain.”

Two new academic papers boast about what they've been up to in more detail. In one, the team explain how they've mitigated "false trigger mechanism" events where users call the chatbot by mistake and where the resulting noise creates poor training data. In the other, they explain how they're training voice trigger detectors on speech data, even when speakers have heavy accents using a "novel voice trigger detector that can use a small number of utterances from a target speaker to improve detection accuracy."  

California-based technology firms with engineers working on chatbots may want to lock-in their staff. Chandra didn't respond to a request to comment on his future hiring plans, but it's understood that the team is always open to hiring opportunistically.