Efficient local inference
How compact models can deliver useful capability within real memory, power, and latency limits.
Agxiom / Local intelligence company
Intelligence,
closer to the metal.
We research local language models and turn the useful findings into private, resilient software.
Research / Local intelligence
Small models. Serious questions.
We study what happens when language models leave the datacentre and meet real devices, private data, limited memory, and imperfect networks.
Behaviour under constraint
How smaller models reason, fail, and change when the environment is private, offline, or resource-bound.
Useful systems
How runtimes, interfaces, and evaluation make local models dependable enough for practical software.
Local intelligence stack
One system, all the way down.
Local AI is not a model dropped into an application. Capability, inference, hardware, and the data boundary have to be designed as one system.
Compact models, deliberate evaluation, and behaviour that remains understandable under constraint.
Memory, latency, quantisation, and orchestration treated as product decisions rather than afterthoughts.
Private and resilient operation where data is created, including offline and resource-bound environments.
Our products
Research becomes software.
Agxiom is building a portfolio of local-first tools and products. The first releases are being shaped now.
Private model runtime
A focused path from compact model to dependable local operation.
- Compact-model path
- Private operation
- Hardware-aware
Local research toolkit
Tools for observing, comparing, and improving models under real constraints.
- Model behaviour
- Runtime traces
- Real constraints
Applied intelligence software
Products designed around useful private intelligence rather than an API dependency.
- Private workflows
- Resilient operation
- Useful interfaces
About us / Agxiom
Theory meets the real machine.
We bring physics and software engineering to the same question: how can language models become useful, dependable systems that run close to the people using them?
Independent by design
Agxiom connects an interest in how models work with the practical work of making them run. Our focus is local intelligence: understanding model behaviour, working within hardware limits, and building software around what we learn.
Start a conversationSoftware / Applied AI
George Liu
From an experiment to a system people can use.
George builds enterprise integrations, platform features, and developer tools at Versori, part of Avalara. His work spans backend and frontend engineering, AI-assisted retrieval, and the realities of running software in production.
Alongside that work, he experiments with local language models and agents, exploring quantisation, GPU and CPU offloading, KV-cache optimisation, and inference profiling. He holds a BSc in Computer Science from Swansea University.
- Background
- Computer science & production software
- Exploring
- Local inference & agent systems
Physics / LLM theory
Arian
An interest in the theory behind language models.
Arian previously studied at the University of Cambridge and is now a master’s student in physics at the University of Manchester. He has a particular interest in the theory of large language models.
His physics background brings a theoretical perspective to Agxiom’s work on local intelligence, alongside the practical questions of how models behave and run on real hardware.
- Background
- Cambridge → Manchester · Physics
- Exploring
- Large language model theory
How we approach the work
Local by design
Privacy, resilience, latency, and ownership are architectural inputs—not features added later.
Constraints are inputs
Memory, power, hardware, and imperfect networks reveal what a useful system actually needs to be.
Research becomes product
We carry evidence through the whole path, from experiment to runtime to software people can use.
Systems should endure
The result should remain understandable, maintainable, and valuable when the novelty wears off.
Contact / Direct line
Bring us the problem that needs to run locally.
We are interested in research collaborations, applied local-AI problems, and the software questions that sit between model and machine.