Custodia has launched Sentinel, a personal AI device for executives, researchers and family offices. The Swiss company says the appliance keeps data off the internet by running locally.
The standalone hardware unit is designed to analyse documents without sending queries or files to external servers. Users choose which files the system can access, and the device operates separately from a customer's existing computers and networks.
Sentinel enters a market where most widely used AI tools rely on cloud infrastructure run by large technology groups. Custodia is targeting customers in sectors such as life sciences, finance and advisory work, where research, investment plans and client records are commercially sensitive.
Working units are already being used by researchers, advisers and executives, according to the company. Custodia describes the device as a one-off purchase rather than a subscription service, with no usage credits or per-query charges.
Local model
At the centre of the system is a retrieval-augmented generation setup, which draws answers from a defined set of user-supplied documents rather than from open internet sources. Custodia says this is intended to reduce the risk of responses being shaped by irrelevant material and to keep the model focused on the information loaded onto the device.
The design reflects a broader debate over data control in AI. Many businesses have tested generative AI tools for internal research and drafting, but concerns about privacy, intellectual property and legal exposure have led some organisations to restrict the use of public systems for sensitive work.
Tom Brooks, Chief Executive Officer of Inhalis Therapeutics and Co-Founder of Custodia, described using the product with pharmaceutical data.
"At Inhalis Therapeutics, we rely on the continuous assessment of new data from studies in the fast-changing therapeutics landscape. I had the perfect opportunity to test Sentinel. I loaded the raw data from our most recent pharmacokinetics study, as well as a precise set of scientific publications describing disease processes, and queried Sentinel for the rationale of a strategic collaboration. This seems like a perfect task for an LLM, but it would be irresponsible for us to dialogue with remote AI systems, regardless of their promises to protect and encrypt the data. The risk is too high and the context too vast," Brooks said.
He gave further detail on how the system was used in an isolated environment.
"Sentinel has allowed us to build a safe, air-gapped 'walled garden' containing only the data we intend within exactly the context we describe. The results were instantaneous; I could rely on the privacy of my conversation with Sentinel as well as the focused scope of its reasoning. This kind of strategic reasoning would have been impossible without Sentinel," Brooks said.
Target users
Custodia is pitching Sentinel to a narrow group of buyers rather than the broader consumer AI market. It identifies family offices, pharmaceutical researchers, legal advisers, financial professionals and senior executives as its main users, arguing that they need tighter control over where data is stored and processed.
One example cited by the company involved a family office using the system to preserve a founder's accumulated knowledge and decision-making record. In that case, the device was trained on selected internal material so staff could query it in a way that reflected past strategy and judgment.
Custodia argues that this kind of product is as much about jurisdiction as cyber security. By keeping the model and data on a local device, customers avoid reliance on third-party cloud providers and the legal frameworks that govern data held on overseas servers, it says.
Company background
Custodia is based in Lugano and has focused on AI data governance and local AI systems for regulated industries. According to the company, Sentinel is its main commercial product and grew out of that work after clients sought private AI tools that did not depend on cloud services.
Federico Brooks, the company's Co-Founder and Chief Technology Officer, built the device. According to the company's description of its leadership team, his background includes product management and software engineering work related to AI and data privacy.
Sentinel is available to a limited group of early adopters through a waiting list, with units configured individually, the company says. Custodia did not disclose pricing, but says the product is aimed at professional and institutional users rather than the mass market.
The commercial challenge for Custodia will be persuading cautious buyers that a dedicated AI appliance justifies the cost and operational effort of installing separate hardware. Its opportunity lies in a corner of the market where convenience is often secondary to confidentiality, ownership and control over sensitive information.
Custodia says neither it nor third parties can access a customer's device, read the data on it or make changes without the user's involvement.