A research laboratory devoted to AI for Science algorithms and their applications; to the architecture and management of advanced AI & HPC clusters; and to the continuous improvement of AI & HPC application and model efficiency and performance.
We design and build the algorithmic, architectural, and engineering foundations that allow artificial intelligence to serve as a rigorous instrument of scientific discovery — and high-performance computing to serve as a trustworthy substrate beneath it.
IdeaHorizon is built on a single conviction: the next generation of scientific breakthroughs will not come from a single discipline, but from the disciplined convergence of artificial intelligence, high-performance computing, and the engineering that makes both usable.
We work on three intertwined vectors — the algorithms that let AI reason about the physical world, the cluster architectures and management stacks that let those algorithms run at scale, and the systems-level optimisations that close the last mile between a researcher's question and a trustworthy answer. Each line of research is shaped by, and shapes, the others.
Our outputs are open standards, peer-reviewed work, and two production systems that ship in real research environments: ScienceMate, an end-to-end agentic platform for scientific research, and PODsys, a software stack that turns bare metal into a production-grade AI cluster in minutes.
We organise our work around three mutually reinforcing vectors. Progress in any one of them tightens the loop on the other two.
Algorithms & Applications
Agent systems and learning algorithms purpose-built for the way science actually gets done — hypothesis, experiment, falsification, reproduction — rather than the way chatbots hold a conversation. Applied across computational simulation, literature research, and formal derivation.
Cluster Design & Management
Hardware-aware cluster architecture, unified control planes, and the management software that lets large-scale AI and HPC infrastructures behave as a single, observable, governable substrate — across on-premise and hybrid deployments.
Systems Optimisation
The last-mile engineering that decides whether a research idea is actually usable: parallel runtimes, kernel- and driver-level tuning, network and storage optimisation, model-system co-design, and the measurement infrastructure that keeps all of it honest.
A digital research group, scheduled by AI, fully auditable, and reproducible end to end. ScienceMate is not a chatbot with extra tools; it is a complete agentic platform built around the way science actually works — hypothesis, experiment, falsification, accumulation.
“
Return the scientist to science · leave the workflow to the platform · keep the judgement with the human.
— ScienceMate Design Principle
From bare metal to a production-grade AI cluster in twenty minutes. PODsys is a software stack that handles the entire lifecycle of a large AI / HPC cluster — installation, identity, scheduling, monitoring, and continuous performance optimisation.
We publish both as papers and as running systems. A complete publication list is being assembled and will appear here as work is released.
Architecture, design principles, and production evaluation of an Orchestrator-coordinated platform of specialised nodes for end-to-end scientific research.
For research collaboration, system access, academic enquiries, or press, please use the channels below.