NPCI And Nvidia To Build Sovereign AI Layer For Payments
FiMI was developed in-house by NPCI to address the limitations of general-purpose language models in high-scale, high-trust payment settings and it is currently deployed to power the Unified Payments Interface (UPI) Help Assistant pilot at national scale. The model was trained and fine tuned on Indian financial and synthetically generated payments data to support tasks such as transaction dispute resolution, mandate lifecycle management and responses to regulatory or ecosystem queries. A technical paper describing its construction and evaluation has been published by the organisation.
NPCI said it plans to incorporate Nvidia's NeMo framework and Nemotron family of models into its development pipeline, exploring the Nemotron 3 Nano model alongside proprietary datasets to advance multilingual customer service across the banking ecosystem. The next phase is described as moving from use case specific agents to a generalised, scalable AI layer, with consideration of architectures such as Mixture of Experts to meet high-volume, low-latency requirements and to broaden agent oriented capabilities.
The organisation characterised the in-house approach as rooted in trust and customer experience, with an AI strategy built around three pillars: internal efficiency, reusable models for the wider payments ecosystem, and systems that reflect Indian data, languages and context. The UPI Help Assistant supports English, Hindi, Telugu and Bengali with more Indian languages expected within six to eight months, and NPCI signalled its intent to enable responsible innovation through governance frameworks and secure infrastructure.