Most public conversations about artificial intelligence focus on the models themselves. Which one is largest? Which one scores highest? Which company released the latest version? The assumption is that progress in AI comes mainly from building bigger, more capable models.
Akhil Pamarthy sees a different source of value. As founder of Offline Intelligence, he is building the infrastructure that surrounds AI models: the systems that decide how a model operates, which tools it uses and how it performs on real hardware in real environments. He believes these systems matter just as much as the model, and in some cases more.
“At Offline Intelligence, we build infrastructure and indigenous, autonomous systems and software that harness the intelligence and capabilities of AI,” Pamarthy says.
Same Model, Different Results
A recent result from the open-source AI community helps illustrate his point. Hugging Face co-founder and CEO Clément Delangue shared that his team had turned several popular AI coding harnesses into reinforcement learning environments. Coding harnesses are the software layers that sit around a model and direct how it works. When the same model, with the same weights, was run inside different harnesses, the results differed sharply: 62 percent under one harness and 33 percent under another.
For Pamarthy, that gap captures the core idea behind his company.
“What changed between 62 percent and 33 percent is not the model, not the task, but the harnessing systems around the model, which decide what tools to call and how to operate,” he says.
If the systems around a model can nearly double its performance on the same task, then building those systems well is a major opportunity. For organizations that need AI to run privately on their own hardware, it may also be the most important one.
The Foundation: A Runtime Built for No Network
Offline Intelligence’s technology starts at the lowest level. To run AI entirely on an organization’s own machines, with no connection to outside servers, Pamarthy needed an inference runtime designed specifically for that purpose.
He built it himself. In March 2026, he published the Offline Intelligence Library, a cross-platform local inference runtime written in Rust. Rust is a systems programming language valued for its speed, memory safety and reliability, which are important qualities for software expected to run on varied hardware without supervision. The library includes bindings for Python, JavaScript, Java and C++, so developers working in a wide range of environments can build on it.
The runtime is the base layer for everything else the company builds. It allows AI models to run directly on local hardware, with data staying exactly where the organization keeps it.
Designed for Hardware Limits
Pamarthy approaches this work as both a mechanical engineer and a software engineer. He holds a bachelor’s degree in mechanical engineering and a master’s degree in computer and information sciences, and his engineering background shapes how he designs software.
Cloud AI can draw on vast data centers. AI running on a law firm’s workstation, a hospital server or a field device cannot. It has to perform within the limits of the machine in front of it. Offline Intelligence is designed with that reality in mind, aiming to deliver capable AI with modest compute and energy requirements.
That efficiency is part of the company’s broader promise. Pamarthy wants organizations to own their intelligence and run it on their own infrastructure, without the costs and dependencies that come with renting access to someone else’s.
Air-Gapped From Day One
A defining feature of Offline Intelligence’s architecture is that it was air-gapped from the start. Many AI products are built for the cloud and later adapted for private deployment. Offline Intelligence took the opposite route, designing for complete disconnection first.
That choice has practical benefits. Systems that do not depend on a network cannot be interrupted by one. They do not send data out for processing, and they continue working in environments with limited or no connectivity.
Pamarthy also emphasizes transparency. The company’s technology is designed to be transparent and auditable, which matters for organizations that must understand, verify and justify the systems they rely on.
And rather than focusing only on inference, generating answers from a model, Offline Intelligence is focused on performing operations in real time, supporting decisions and actions as they happen.
From Law Offices to Mission-Critical Operations
The company applies this foundation through two main products.
Offline Counsel AI is a private legal AI built for law firms, in-house legal teams and legal practitioners, available on Windows and macOS. It is designed for work such as legal research, workflow automation, offline e-discovery, privileged document redaction, due diligence and secure contract analysis, all without client files leaving the firm’s own environment.
Aurora OS takes the same principles into far more demanding settings. It is designed to deliver battlefield intelligence for mission-critical operations, where systems must perform under pressure and often without any network at all.
The company also builds custom AI infrastructure and software for organizations with specific data-sensitive requirements.
Offline Intelligence’s core technology, operating system and legal software have been built, tested and validated. The company is now working toward pilot customers and proofs of concept ahead of its planned launch in November 2026.
Building Toward Hardware
Pamarthy’s technical roadmap moves steadily closer to the machine. After the application layer, the company plans to expand into hardware and compute. That includes hardened, on-premises AI appliances, such as rack-mounted servers and local workstations, for organizations that want data to never leave the building.
It also includes low size, weight and power edge devices, compact systems capable of running AI in contested, disconnected environments such as drones, naval vessels and forward operating bases.
Longer term, Pamarthy plans to develop the company’s own small language model. He is clear about the order, however. The first priority is the harnessing technology around models, where he believes the most immediate value lies.
An Engineering Team in the Making
So far, much of the core technology has been built by Pamarthy himself. As the company prepares to launch, he is focused on growing its engineering capacity, hiring engineering and sales talent and searching for co-founders who share his vision. Offline Intelligence plans to share more about its team and research as that work progresses.
Intelligence That Belongs to Its Owner
Behind all of the technology is a single goal. Pamarthy wants every organization to be able to run its own private, self-contained intelligence, on its own desktops, servers, edge devices, rugged hardware, IoT systems and robotics.
The models will keep improving. But Pamarthy is betting that the companies shaping the next era of AI will be the ones that build the systems that let those models run reliably, privately and independently, wherever the work happens.
