Research and applied AI
Language models that know their limits.
XenAI is an independent research lab for applied AI. We study how to control the risk of language model systems that support high-stakes decisions, beginning with the law.
The failure we study
Right source, wrong date.
A language model can cite a real statute, summarise it accurately, and still give the wrong answer, because that law was not in force on the date the question is about. Checking that a citation exists does not catch this.
The question
A couple in England wanted to divorce in 2021 without blaming each other and without first living apart for two years. Could they?
An answer that looks right
Yes. The Divorce, Dissolution and Separation Act 2020 allows a divorce without any allegation of fault.
The right answer
No. That Act came into force on 6 April 2022. In 2021 the couple needed an allegation of fault or at least two years of separation.
Confidence, claim by claim.
An answer is a set of separate claims. Each should carry its own measure of confidence, so that one weak claim cannot hide inside a fluent paragraph.
Evidence valid on the date.
A source has to be real, relevant and in force at the time in question. We treat the third test as seriously as the first two.
Escalate, or decline.
When a small model is not sure enough, the question should go to a stronger check, a larger model or a person. Each step costs time, money or data leaving the building. Sometimes the right output is no answer.
What we do
Research
Risk-controlled language model systems for high-stakes decision support. Our work covers claim-level calibration, temporally valid evidence and resource-adaptive escalation. Experiments and results are published in Bench Notes as they are ready.
Read Bench NotesApplied AI
XenAI Applied AI turns that research into working systems for legal and educational settings, and advises teams building their own. Family law in England and Wales is our first case study.
About Applied AIAcademy
XenAI Academy trains engineering teams and individuals to run language models themselves: AI platform engineering, LLMOps and SLMOps, and serving models in production with FastAPI.
See the Academy