What is the story about?
The National Payments Corporation of India (NPCI) has launched an open reinforcement learning (RL) environment for banking AI agents in collaboration with NVIDIA at the Global Fintech Fest (GFF) 2026.
The environment is designed to allow developers, banks, fintechs and research teams to train and benchmark AI agents on Indian banking tasks using synthetic data, without relying on real customer information. NPCI said the approach preserves data sovereignty while supporting the development of agentic AI.
The system enables AI models to learn from multi-turn banking conversations, with performance assessed based on outcomes. It is designed to help conversational assistants evolve into tool-using agents that can resolve customer requests within defined rules.
Developers and institutions can use the environment to train different AI models on common banking tasks and compare their performance before deployment. NPCI said the framework is intended to provide a stable, reusable and cost-effective environment for developing and evaluating banking AI agents.
The environment has been built using NVIDIA NeMo, NVIDIA’s NeMo RL framework, and is designed for contribution to NeMo Gym, an open library for evaluating and improving AI agents through environment-based training.
It also follows open standards, including the Model Context Protocol (MCP), allowing institutions to train and benchmark AI models on Indian banking tasks.
The environment can be extended by developers and institutions to cover additional banking tasks and domains. Participants can also contribute new tasks, tools and improvements, allowing it to expand to other banking use cases.
NPCI said the initiative extends its digital public infrastructure approach to AI training by creating an open and governed foundation for the ecosystem. By enabling training, evaluation and comparison on common tasks, the collaboration aims to establish a shared standard for banking AI agents in India.
Also read: Jensen Huang says AGI has arrived: What is GPT-6 Astra, how it works and price
The environment is designed to allow developers, banks, fintechs and research teams to train and benchmark AI agents on Indian banking tasks using synthetic data, without relying on real customer information. NPCI said the approach preserves data sovereignty while supporting the development of agentic AI.
The system enables AI models to learn from multi-turn banking conversations, with performance assessed based on outcomes. It is designed to help conversational assistants evolve into tool-using agents that can resolve customer requests within defined rules.
Developers and institutions can use the environment to train different AI models on common banking tasks and compare their performance before deployment. NPCI said the framework is intended to provide a stable, reusable and cost-effective environment for developing and evaluating banking AI agents.
The environment has been built using NVIDIA NeMo, NVIDIA’s NeMo RL framework, and is designed for contribution to NeMo Gym, an open library for evaluating and improving AI agents through environment-based training.
It also follows open standards, including the Model Context Protocol (MCP), allowing institutions to train and benchmark AI models on Indian banking tasks.
The environment can be extended by developers and institutions to cover additional banking tasks and domains. Participants can also contribute new tasks, tools and improvements, allowing it to expand to other banking use cases.
NPCI said the initiative extends its digital public infrastructure approach to AI training by creating an open and governed foundation for the ecosystem. By enabling training, evaluation and comparison on common tasks, the collaboration aims to establish a shared standard for banking AI agents in India.
Also read: Jensen Huang says AGI has arrived: What is GPT-6 Astra, how it works and price
/images/ppid_59c68470-image-178903752120483318.webp)




/images/ppid_59c68470-image-178903755340521786.webp)
/images/ppid_59c68470-image-178902005702641873.webp)

/images/ppid_59c68470-image-178893752656920849.webp)

/images/ppid_59c68470-image-17889225212791056.webp)
/images/ppid_59c68470-image-178879003160795863.webp)
/images/ppid_59c68470-image-178893252569524586.webp)
/images/ppid_59c68470-image-178893759854217065.webp)