AI research you can check, built for work that matters.
Oxtractor is a boutique AI research and development company in Oxford. We design multi-agent retrieval and conversational AI systems for medicine, science, education and regulated industry, and we take them from the laboratory into practice with university and industry partners.
Oxford University Innovation portfolio companyResearch partner: UCL Computer ScienceDeployed on premise, with full data sovereignty
Approach
How we work
We are a small team by design. Every system we build starts as a research question, is tested against evidence, and is shaped with the people who will rely on it.
Grounded
Answers are built from retrieved sources, cited back to the document, table or figure they came from.
Checked
Independent judge and compliance agents review each answer before a person sees it, and we measure how often they catch errors.
Sovereign
Sensitive institutions keep their data. Our systems run on premise or in sovereign infrastructure and work with any foundation model.
Collaborative
We work alongside universities, clinicians and domain experts, and publish what we learn.
Research
The AI Scientist
Our current programme, run with MSc researchers at UCL's Department of Computer Science, asks what an AI agent must prove before a laboratory relies on it. We frame the answer as three tests.
Test one
Reading
Can it find and use the right evidence?
Agents that search the literature, call scientific tools and choose how to retrieve from long documents.
Tool-augmented research agents
Retrieval configuration for long contexts
Multimodal retrieval over tables and figures
Test two
Trusting
Can its claims be checked?
Methods that locate unsupported statements, audit reasoning claim by claim and trace how agents reached an answer.
Evaluation metrics for RAG
Claim-level auditing
Multi-agent trace auditing
Test three
Sharing
Can institutions improve it together?
Ways for hospitals and laboratories to fine-tune a shared model without pooling their data.
Federated fine-tuning
Reproducibility under uneven data
Robustness when partners drop out
Where we come from
Oxtractor grew out of information extraction research at the University of Oxford's Department of Computer Science, including the ERC-funded DIADEM project on automated web data extraction and our own work on structured aspect extraction.
Today that lineage continues in retrieval, reasoning and verification for large language model systems.
Platform
Oxtractor Nucleus
One research platform underneath everything we build. It is independent of any single industry or foundation model: domain knowledge arrives as packs, and the same verified core serves each partner.
Domain packs
MedicineScientific researchEducationFinance and law
Our multi-agent retrieval architecture. A router chooses the retrieval strategy, a reasoning agent assembles the answer, and compliance and judge agents check it before it is returned.
Product
Deep Conversational AI
Conversational AI and tutoring for learning and research, co-designed with clinicians and educators, with a current focus on medical and scientific training.
Evaluation
Verification suite
Benchmarks and auditing tools that measure grounding, error detection and consistency, so partners know a system's limits before it goes live.
Domains
Where we focus
We choose fields where answers must be traceable and mistakes are costly.
Healthcare and life sciences
Clinical knowledge support, multi-hop medical question answering and literature-aware assistants, built with practising clinicians.
Primary care
Rehabilitation medicine
Reproductive medicine
Scientific research
Agents that read the literature, generate and test hypotheses, and report what they found with the evidence attached.
Literature review
Experiment planning
Result verification
Education
AI tutoring, curriculum mapping and AI literacy programmes for universities and medical schools.
Tutoring dialogue
Curriculum mapping
AI literacy
Finance and law
Document review and compliance checking where every conclusion has to cite its rule and its source.
Trade finance compliance
Regulatory review
Legal research
Partnering
Ways to work with us
Oxtractor is a research and development company. We do not sell subscriptions; we build systems and evidence together with partners.
Joint R&D
A shared research programme on a problem your organisation owns, with clear milestones and published results where possible.
Research consortia
We join and lead grant-funded consortia with universities, hospitals and industry in the UK and Europe.
Applied programmes
We adapt Nucleus to a partner's domain, data and infrastructure, and deploy it inside their walls.
Licensing
Our architectures, evaluation methods and domain packs are available to license for partners who build in house.
Partners
In good company
Our work is anchored in two of the UK's leading computer science communities.
Portfolio partner
Oxford University Innovation
The University of Oxford's research commercialisation company. Oxtractor is part of its portfolio.