Interspatial OS

Digital infrastructure in modern urban environments is currently very Internet-centric, and involves transmitting data to physically remote environments. The cost for this is data insecurity, high response latency and unpredictable reliability of services. I am working on Osmose -- a new OS architecture that inverts the current model by building an operating system designed to securely connect physical spaces with extremely low latency, high bandwidth local-area computation capabilities and service discovery.

In 2018, I was starting to wrap up a multi-year focus on Unikernels, and I went back to look over the state of personal data handling (as I'd finished working on Personal Containers in 2016). Things had regressed fairly dramatically -- central cloud providers and particularly IoT manufacturers were moving heavily towards ubiquitous surveillance and centralised management.

I started with trying to find a different slant on existing architectures for smart buildings. Why couldn't we invert the Internet so that data is pooled in a single physical location by default, with networking being opt-in? Why can't we build all of our ubiquitous computing infrastructure (such as voice and face recognition) so that it runs locally within the building rather than streamed from remote datacentres? There would be gains all around -- latency, energy usage, offline operation -- if we could figure out how to deploy local machine learning services.

I wrote up the initial thoughts behind this in a workshop paper in An architecture for interspatial communication. Since then, I've been collaborating with the good folks at Tarides on building out the library infrastructure in MirageOS to setup a prototype set of rooms in Cambridge and Paris that can act as a testbed for our ideas.

The intention behind the Osmose design is to "invert" the architecture of smart cities to be self-contained units by default, and only communicate when required for the purpose of remote interaction. All sensing and storage is conducted locally -- resulting in energy efficiency gains, security by default for sensitive data, and robustness against communications outages affecting critical physical infrastructure.

Two significants advances in 2023 and 2024 on this project were:

Where on Earth is the Spatial Name System? which explores a DNS architecture for naming places
Scheduling for Reduced Tail Task Latencies in Highly Utilized Datacenters which explores a decentralised scheduling architecture for lower job completion times

Ultra-Low-Power AI Infrastructure

A significant development in 2024-2025 has been our work on ultra-low-power neural processing units for edge deployment, directly supporting the Osmose vision of local AI services. Our benchmarking work provides the first comparative evaluation of commercially-available μNPUs, revealing surprising disparities between hardware specifications and actual performance.

This connects to our broader research on energy-aware deep learning for resource-constrained hardware. The combination of energy harvesting, intermittent operation, and sophisticated AI processing represents exactly the kind of intersection we need for sustainable smart building infrastructure.