As regulatory compliance and data security concerns take centre stage, conversations around AI and GenAI are moving towards a more transparent open source system. According to the World Economic Forum, open source AI development offers three key advantages. It accelerates innovation while maintaining product safety, makes models more reliable, as datasets and codes are thoroughly audited, and significantly lowers time to market and development costs.
Conversely, closed-source models such as the PaLM (Pathways Language Model, a family of large language models developed by Google), most of the OpenAI GPT models, or Claude, an advanced language model by Anthropic, evade audits because these are not publicly available for developers to use. Besides, the data in use is organisation-specific, which is a practice that raises questions about the responsible use of AI.
India might have missed the bus when the open source initiative (OSI) first made a foray in the late 1990s and the first decade of the new millennium. But today, tech entrepreneurs have started building open source AI models and platforms for new applications and better data security. It is a timely push towards ‘Sovereign AI’, boosting the country’s ability to develop, deploy and control AI-related intellectual properties in tune with its priorities.
Take, for instance, NimbleEdge, a San Francisco- and Bengaluru-based AI startup set up in 2021 by AI researcher Varun Khare. It has recently launched DeliteAI, a full-stack and production-ready on-device agentic AI platform, enabling developers to build, deploy and run scalable and privacy-first AI-native solutions directly on smartphones. Additionally, there is a dedicated marketplace to find and integrate pre-built AI agents with applications. NimbleEdge has also introduced an on-device conversational AI assistant with built-in productivity features (more on its technology later).
In an AI era dominated by cloud-based LLMs (large language models) and big tech companies wresting control of global user data, NimbleEdge is making a contrarian play by bringing AI model development to consumer devices to empower companies, developers and users. It is a move away from the mainframe (think of the massive shift from desktops to laptops and handhelds) for better operations, greater control and stringent privacy levels.
Ditching the traditional cloud infrastructure is a case in point. It counters the limitations of cloud-based AI, which requires constant connectivity, increases latency and incurs high operational costs. Furthermore, developers will own the entire stack — everything from models to intelligence integration — enabling unique, AI-powered user experiences. Better still, when data is processed locally on the hardware in use, it enhances data security and user privacy.
“AI is the electricity of our generation. When we remove all the hype, it is still a fundamental shift in technology. This essentially means three things: Scaling beyond what the current cloud can do because the cloud infrastructure in India is very limited compared to the smartphone infrastructure. Again, AI must be personal and adapt quickly as the use of technology varies vastly among different generations. More importantly, it has to be private and secure,” said cofounder and CTO Neeraj Poddar, who joined Khare in 2024.
Why Edge AI Forms The Core Of NimbleEdge
Before we delve deeper into the tech components of DeliteAI and their benefits, it is essential to understand the tech philosophy behind shifting the mainframe, introduced by NimbleEdge. Unlike traditional AI, which relies on cloud-based backend systems to process data, Edge AI is all about building AI algorithms and models directly on connected devices located at the ‘edge’ of the network, closer to the data source, instead of relying on central cloud servers or remote data centres for processing.
This is, undoubtedly, a paradigm shift, as computation happens on the hardware itself (smartphones or other IoT devices) or a nearby Edge server. This decentralised approach allows devices to collect data, analyse it and make real-time decisions using built-in machine learning (ML) models.
Simply put, NimbleEdge is working at the intersection of data technology, artificial intelligence, and the collective intelligence of billions of tech enthusiasts keen to build personalised AI models. In doing so, it is eliminating the traditional dependency on cloud or high-end GPUs (graphic processing units) and ushering in a whole new architecture aligned with the power of smartphones.
Khare seems quite confident about the outcome. Elsewhere, in an interaction, he said that NimbleEdge could run 100 Mn AI models personalised for 100 Mn users.
But first things first. Khare, a graduate in computer science from IIT Kanpur, joined the University of California at Berkeley in 2020 for his research and started working with Professor Dawn Song, an AI and privacy advisor to the US Congress. Under her guidance, he realised the profound impact of AI when placed in human hands. Soon enough, he started leading the on-device training team at OpenMined, one of the biggest communities developing privacy-first infrastructure around AI and ML.
The startup’s first major offering was an SDK (software development kit) for AI-powered app development. Within the first two years, its early versions reached more than 30 Mn devices, as NimbleEdge partnered with some of the biggest ecommerce and fantasy sports platforms. However, it declined to reveal customers’ names, citing non-disclosure agreements.
Last year, Khare was joined by Neeraj Poddar, a seasoned entrepreneur with a decade of experience. Poddar’s first venture was Aspen Mesh, alongside an open source community that he built for about five years. Next, the duo focussed on two key areas — making the SDK an open source platform and launching a dedicated agent marketplace for developers.
Their efforts paid off, and DeliteAI was globally launched on July 10, 2025.
On the B2C front, NimbleEdge has developed an AI Assistant, which it claims to be an industry-first, fully on-device conversational AI assistant with built-in productivity capabilities. open source platform. In the past four years, the startup has partnered with Swiggy, SarvamAI, Meta, Qualcomm and several other industry leaders to develop its open source platform.
A Deep Dive Into The Open Source Stack
As discussed, the latest tech stack driving the DeliteAI platform includes a production-ready, open source SDK, an agent marketplace and the NimbleEdge AI voice assistant for end users.
To begin with, the platform is natively built for smartphones (both Android and iOS), addresses many hardware constraints and bridges critical gaps such as the lack of unified tooling and standardised runtimes across the on-device AI ecosystem.
Its SDK is powered by an optimised inference engine, responsible for reasoning and decision-making, and an on-device Python runtime that manages agentic workflows using Python scripts. These components integrate seamlessly with industry-standard runtimes like ONNX and ExecuTorch, abstract away different mobile hardware and enable developers to create dynamic applications using familiar tools. The open source stack also ensures quick cross-platform deployment of LLMs, multimodal AI and transformer models without GPUs or cloud infrastructure.
On DeliteAI, workflows and models can be updated on the fly via a SaaS tool without releasing new apps. Again, user interactions are captured in real time, and intent-driven agents are activated for diverse use cases, from ecommerce to entertainment and grocery shopping.
According to Poddar, no one has been able to run Python on phones until now, while integrating and orchestrating AI on phones remains a significant challenge.
“Besides, our open source SDK will unlock a whole new experience for developers to enhance their applications. They can connect our on-device SDK to our SaaS platform, or link their custom SaaS platforms to update/change models or Python scripts. This will be a game-changer for our customers to explore innovation faster,” the CTO added.
As DeliteAI is built for complete customisation and developer control, companies and individuals can bring their proprietary models and run inference directly on the device for enterprise-grade customisation. Processing everything locally ensures no data leaves the device, allowing organisations to fine-tune and deploy large language models and agents entirely offline.
The Agent Marketplace at NimbleEdge provides a fast-growing library of plug-and-play AI agents for tasks like summarisation, recommendations and speech processing.
Finally, there is the on-device, conversational AI assistant with built-in productivity use cases. But unlike proprietary assistants bound to a single ecosystem, NimbleEdge enables fully customised workflows and branded assistants, all powered by an open source, on-device platform without relying on external dependencies.
However, the startup faces stiff competition from global players like Apple’s Siri, Google Assistant, and a growing list of email summarisation and note-taking platforms like Fireflies and Fathom.
Khare argues that Siri and Google Assistant struggle to scale as they require access to multiple applications, while many users are wary of sharing data.
“In contrast, NimbleEdge integrates seamlessly with Slack, Gmail, phone voice recorders and other apps, without requiring users to send their data to third-party developers,” he explained.
Can NimbleEdge Compete With Global Giants, Align With India’s AI Mission?
NimbleEdge has started generating revenue but has not disclosed its financials. But going forward, it will place its bets on a B2B subscription model as its primary revenue stream. Currently, customers are charged based on the number of devices using the application. For smaller apps with around 10K monthly active users, the cost per device is 20 cents per day. However, the pricing will change with the increasing integration of LLMs into these applications.
Within the next 12 months, NimbleEdge expects to reach 100 Mn devices and a subsequent jump in revenue. But in spite of this business growth, its core focus will be expanding its open source community.
There lies the crux. Unlocking the full potential of AI through on-device, open-source operations and scaling personalised models to deliver engaging experiences should be considered a leap forward. As Podder highlighted in a recent press statement: This is the missing infrastructure layer and developer tooling we longed for while building distributed systems at a global scale. Now, it’s accessible to everyone.
Still, can it scale the open-source business fast enough to compete globally?
The AI industry remains a fiercely competitive, proprietary space, where the adoption of open source solutions is far from widespread. Unlike China, where the OpenAtom Foundation — a non-profit formed in 2020 with support from major corporations and the Ministry of Civil Affairs — has been a catalyst for large-scale open-source AI projects, India has yet to establish a comprehensive policy.
The rise of big tech companies like Microsoft, Meta, and Google in this space is another cause for concern. For instance, Microsoft created Open Neural Network Exchange (ONNX), and Meta developed PyTorch, a popular open-source machine learning library widely leveraged for R&D and product development.
NimbleEdge is not overly worried, though, saying it isn’t in direct competition with these tech giants. In fact, its platform is built on top of these frameworks. “Our objective is to efficiently run open source models like Llama (from Meta) on smartphones. We are also training some predictive models,” said founder-CEO Khare.
Although the startup intends to develop proprietary models in the future, the founders kept the details under wraps. “Building this ecosystem starts with open source technology. We have already crafted a solid monetisation strategy and built a SaaS platform, which delivers new configurations and updates, tracks model performance, and improves user experience. We believe this business model will thrive,” said Poodar.
Furthermore, the startup’s decentralised, privacy-first approach and data localisation policy may align with India’s sovereign AI mission. This will lead to new opportunities to contribute to a trusted, open source AI infrastructure.
Given its primary focus on the US and India markets, it will be interesting to watch how NimbleEdge spearheads its efforts to shape a safer, more regulated AI landscape locally and globally.
[Edited by Sanghamitra Mandal]
