Artificial Real Intelligence
Eigenform builds local, empirically grounded recursive self-improvement architectures. Our systems formulate hypotheses about their world, write code to test them, evaluate results, and retrain on what they've learned, just like a human researcher. Each cycle produces a verifiable reasoning trace.
Reach out to usMachines that learn from experience
Eigenform builds local artificial intelligence that can act, learn from what happens, and use that experience to improve the model itself. We build full-stack local intelligence: from secure durable local compute boxes, to adaptable harnesses, to models that manage their own retraining.
State-of-the-Art performance from secure local models
The Eigenform stack is designed for environments where conventional cloud AI isn't enough: remote operations with unreliable connectivity, organisations that need sensitive data and compute to remain under their own control, and researchers working with specialist or cutting-edge datasets that general-purpose models have never seen. By bringing compute, data and learning into the same local environment, models can work with knowledge that is private, new or unique, and adapt to it over time.
Trusted by organisations including
The Eigenform Self-Improving AI Stack
Self-improving AI requires more than a model. It requires a persistent loop connecting intelligence to an environment in which its ideas can be tested.
Environment
Documents, images, sensors, databases, software and measurable real-world outcomes.
Agent Harness
Models inspect the environment, formulate hypotheses, write code, use tools and take actions.
Empirical Evaluation
The consequences of those actions provide evidence about which behaviours and hypotheses are useful.
Continual Finetuning on Local Hardware
Successful interactions become training data that can inform subsequent model generations.
AI That Lives Where It Learns
Complexity, not price, has traditionally been the major obstacle to local AI inference. We want to make owning a GPU as common and obvious a thing as owning a laptop.
Our prototype combines GPU compute, local storage and a complete computer in a robust, portable form factor, enabling AI inference and data processing without dependence on continuous cloud connectivity.
CAD designs and build photos
Research you can use.
Geocluster
An open-source AI workspace for geological exploration.
Our tech is already being applied in the resources and earth sciences sector worldwide, via the Geocluster platform.
Geocluster is available as an open source product: all processing can be done locally so your discoveries stay yours.
Free forever · Local-first · Bring your own model
Geological and Stratigraphic Data Processing
Clean and structure legacy reports, maps, tables and exploration datasets.
AI-Assisted Exploration Analysis
Let agents inspect underlying data, write analysis code and explain their reasoning.
Physics-Informed 2D and 3D Visualisation and SoTA Map Vectorisation
Explore spatial relationships, geological structures and model outputs directly.
From Raw Data to Models That Improve
- 01
Build the Environment
Reports, maps, drillholes, assays and geophysical data are transformed into machine-usable evidence.
- 02
Generate Geological Hypotheses
Agents investigate relationships within the data and propose falsifiable explanations.
- 03
Test Them
Candidate hypotheses are evaluated programmatically against spatial evidence.
- 04
Keep What Works
Useful hypotheses contribute to a persistent representation of the geological environment; weak or redundant hypotheses are rejected.
- 05
Train the Next Generation
Successful reasoning traces become supervised training data for subsequent model generations.
Does Self-Improving AI Actually Improve?
Negative Space Learning
Learning from consequences instead of labels
Our Negative Space Learning research investigates whether agents can generate useful training data through successful interaction with an external environment rather than relying entirely on human demonstrations or model-generated judgments. NSL trains from physical, measurable consequences of the agent's own actions.
Read the research →Continual Learning in the Field
1,137 experiments · 60 proven relationships · +17.6% model uplift
In the reported geological experiment, successful experimental traces were used to train subsequent generations, producing a 17.6% improvement over the base model on a separate geology benchmark.
See the experiment →Blind Exploration
Could an AI make a discovery using only information from before it happened?
In the "Coe Fairbairn" retrospective test, Eigenform restricted the system to exploration data predating a 2021 rare-earth discovery and tested whether it could identify the subsequently discovered area.
See the experiment →What Makes Us Unique: AI Agents Can Remember. We Want Them to Learn.
Most AI agents can accumulate context, write to memory and use external tools. But the underlying model remains largely unchanged.
Eigenform researches a harder form of continual learning: systems that turn successful interaction with their environment into new training data and ultimately into changes at the weight level.
What we haven't solved: the path to Recursive Superintelligence
A living map of our work — from simple rules to recursive minds
We believe that the future of AI does not just lie in ever bigger models and ever more training data. We believe that as continuous learning and recursive self-improvement become not just more prevalent but increasingly necessary, we will also see a shift towards agile local models that can learn continuously within specific environments in much the same way a human does - something that the bigger models, which must be all things to all users, cannot afford to do. With this in mind, Eigenform develops methods for models to evaluate their own behaviour, learn from the environments in which they operate, and guide their own future development. We study this as a fundamental problem in recursive self-improvement and build practical systems that put the same principles to work today.