About AiirGap
AiirGap finds the statements in an organisation's documents that cannot all be true. It runs on the customer's own hardware and it shows its work.
What AiirGap does
Large organisations contradict themselves. A retention period in one policy disagrees with the one in a procedure written three years later. A specification and its test plan name different limits. A control says one thing and the handbook it cites says another. Each of these is a compliance finding waiting to happen. AiirGap reads the whole set of documents, pairs the passages likely to disagree and classifies whether both can be true, then verifies every evidence quote against its source, character by character.
Why it exists
Document sets grow faster than anyone can read them. Departments, systems and decades each add their own layer. Manual review finds what a reviewer happens to look at. Search finds similar text. Neither finds the pair of sentences, far apart in two different files, that cannot both hold. That gap is what AiirGap was built to close.
How it is built
The product is designed for the audience that will audit it. A single nine-billion-parameter model with specialist adapters runs fully offline on one workstation or on the customer's own GPU server, installed from a sealed bundle, with login, role-based access, two-factor authentication and single sign-on built in. Every prompt, response and probability is stored, so every finding can be attributed and defended later.
The training data went through an adversarial label audit that rejected about a third of the hand-written examples before a model ever saw them. Political neutrality is measured with a standing evaluation that every model change must pass. Energy use is measured too: 3.3 watt-hours per thousand comparisons through the production path. The reasoning behind these choices and the numbers behind them are set out on the homepage and in the research.
Why AiirGap
Documents never leave
Processing happens inside the customer's environment. AiirGap holds no copy of the documents and has no route to them.
Every finding shows its work
Verified evidence quotes, per-token confidence and a readout of the model's internal state accompany each verdict, so a reviewer can check it rather than trust it.
Honest about uncertainty
It says "look at this", not "this is broken". A verification pass filters false positives before a person sees them. The close calls are flagged as close calls.
Built for the audit
Projects, roles, quotas and a full decision log, with exports in CSV and PDF. The audit log has no interactive edit or delete; removal happens only through the logged retention policy.
The founder
AiirGap was founded by Gareth O'Shea, based in Limerick, Ireland. His research is in large language models, in particular open models run locally: how an organisation deploys them on its own hardware so that nothing leaves the building. He speaks regularly on the subject.
He is an expert in contradiction in language: how statements come to disagree and how a large language model can unlock genuinely novel soluutions to find the disagreement and prove it.
He has built two open standards for organisations to aide in the data hygiene problem that breaks AI adoption:
- GASP, the Generally Accepted SaaS Principles, is the open standard for SaaS intelligence: 60 core metrics, 15 entities and 164 relationships that connect them, so that engineers, analysts, operators and AI agents work from one definition of each number. It is published under CC BY 4.0, with code under Apache 2.0.
- The AI Control Framework (AICF) is its companion for governance: 168 canonical controls normalised from eight frameworks, SOC 2, ISO 27001, NIST SP 800-53, NIST AI RMF, the EU AI Act, ISO 42001, CSA CCM and GDPR, with 807 cross-framework mappings and 343 auditable questions. AiirGap is built to the controls it maps, which is why those eight frameworks appear in the footer of every page on this site.
AiirGap applies the same instinct to a organisation's own documents: agree on what is true, write it down once and check that everything else agrees with it. Gareth writes the research published here and can be found on LinkedIn here.