Nvidia and its partner Elite launch nine million synthetic identities to test banking agents
Listen to this article
Read by Anchor
Developing AI agents for the banking sector faces a thick wall of compliance and confidentiality constraints, because testing models on real customer data carries a costly legal risk. To untangle this dilemma, Nvidia, together with Japan's Elite, released an open-source dataset called "NemoTron Personas Finance Japan" that provides nine million synthetic identities reflecting Japan's demographic and financial composition, allowing institutions to train conversational agents and test client behaviour without touching any real banking record.
The new data are based on official government statistics, but have been transformed through probabilistic graph modelling into files that mimic reality with fine detail.The collection consists of one million records, each containing nine distinct personas, covering fifty fields per file, including financial and real-estate assets, income levels, spending habits, digital and financial literacy, as well as more than 1,600 occupational categories and roughly 950,000 unique Japanese names. This output, amounting to one and a half billion tokens, was generated using the Nemo Data Designer system and the probabilistic Nemo SDG-BGM framework, relying on the open Gemini 4 model to produce natural Japanese text, and it is released under the Creative Commons BY 4.0 licence for commercial use.
The project was not detached from operational market needs; it was designed with direct support and feedback from major Tokyo financial institutions, including the Meiji, Sumpo Japan, Tokyo Marine Holdings, Fukuoka Bank and Meiji Yasuda groups, as well as the Financial Data Usage Association. The dataset also integrates directly with the Nemo User Sim user-simulation framework and coincides with the launch of the NemoTron BII Japan package, which contains one hundred thousand synthetic business documents tagged with personal identity data for training refinement systems and masking sensitive information.
The engineering lesson here reshapes the operating rules for developers and CTOs in Gulf, Egyptian and Levantine banks.Regional banks in Riyadh, Abu Dhabi, Cairo and Amman clash with strict central-bank and data-governance regulations that forbid feeding banking transaction records into generative AI training environments, and they impose tight data localisation and supervision. The Japanese experiment shows that the practical solution does not rely on hard-to-obtain legislative exemptions or on sacrificing privacy; it lies in converting public statistical tables and disclosed population distributions into ultra-realistic synthetic environments.
If you lead a technical team at a regional financial institution, your practical step goes beyond the futile pursuit of frozen customer-data licences from compliance committees and moves straight to building probabilistic modelling systems that recreate local household and spending structures in a tightly controlled synthetic format. This path lets you test agents against failure and fraud scenarios, train customer-service tools, and simulate complex credit decisions safely and at a cost that fits sovereign AI priorities.