Applied Research · XEROTECH LTD
Sovereign AI Lab (SAIL)
Regulated organisations cannot use AI the way Silicon Valley does. Their data cannot leave the building. Their compliance obligations do not pause for innovation cycles. The tools being built for them are too expensive, too opaque, or too dependent on infrastructure they do not control. SAIL exists to close that gap.
How it works
Research that ships
Every research track connects directly to a commercial deployment. The chain is deliberate.
SAIL researches it
→
XEROTECH builds it
→
PULVINIR and CallGPT 6X deploy it
The patent-pending privacy filtering in CallGPT 6X came out of SAIL’s privacy-preserving research track. The epistemic confidence framework in PULVINIR — our regulatory intelligence platform for energy, legal, tax, health, and education — came out of the compliance intelligence track. Research here does not sit in a drawer.
Engineering Capabilities
Retrieval-augmented generation over domain-specific knowledge engines
Deterministic rules engines for regulatory penalty modelling and compliance gap scoring
Vector databases for semantic retrieval across regulatory corpora
Multi-tier LLM architectures with audited per-call cost and token telemetry
Automated document generation with epistemic confidence tagging on every claim
Workflow orchestration across cloud and on-premise infrastructure
Client-side data redaction before any information reaches an AI provider
Full audit trail with request-level telemetry across every LLM call
Academic research infrastructure at Artificial Intelligence University. Commercial deployment at XEROTECH.
Published papers
Patents cited by Apple,
Microsoft & Amazon
Regulated sectors
in production
What we research
AI for regulated organisations
The dominant assumption in AI is that deployment requires hyperscale compute and frontier-model APIs. For regulated organisations, that assumption is wrong. SAIL asks what actually works when data cannot leave the building.
Compliance intelligence
Deterministic rules engines produce auditable numbers — penalty bands, downtime costs, maturity scores — without any LLM involvement. AI generates the narrative, grounded in cited sources, with every claim tagged for confidence level. A board, a compliance function, and a regulator can each read the same document and know exactly what is sourced and what is inferred. Built for energy, tax, legal, health, and education sectors.
Small-scale language models
How small can a model be and still do the job? Not as an academic question — as an engineering constraint for organisations that will never rent GPU clusters. Our published ILM model runs inference on hardware architecture from 1999. On a narrowly scoped task, the optimiser choice contributed more measurable improvement than switching from LSTM to Transformer. Useful AI does not require frontier compute. We can prove it with numbers.
Privacy-preserving architecture
AI systems where sensitive data never reaches an external provider. Client-side redaction before any query is transmitted. On-premise inference on hardware the organisation already owns. The patent-pending privacy filtering in CallGPT 6X came directly from this track, as did the deployment patterns used across PULVINIR engagements.
AI governance for regulated environments
What technical and organisational structures make AI safe to deploy when a regulator is watching. Ofgem for energy. FCA and PRA for financial services. CQC for healthcare. SRA for legal. Ofsted and OfS for education. ICO for data protection. NCSC CAF 4.0 for cyber resilience. The EU AI Act across Europe. We study the regulatory instruments, model the enforcement patterns, and publish analysis through the SAIL newsletter.
Published research
Papers, models, and data
Preprints on Zenodo and TechRxiv. Open models and datasets on HuggingFace. Longer-form analysis through AI Business Review.
“On this corpus at this scale, the optimiser contributed more to the measurable improvement than attention did.”
ILM & ArfaLM — SAIL, April 2026
Language models
ILM and ArfaLM
A 2.3M parameter LSTM trained using only techniques available in 1999 versus an 8.4M parameter Transformer with modern training — on identical child-directed speech data. When the performance gap is decomposed, the optimiser change alone accounts for 1.56× of the improvement. The architecture change, tokenisation improvement, and 3.6× parameter increase together account for 1.38×. On a narrowly scoped task at small scale, how you train matters more than what you train. Part 1 published. Part 2 — on-hardware deployment, integer quantisation, Pentium II benchmarks — in progress.
Paper (Zenodo)
·
Models (HuggingFace)
·
Dataset (HuggingFace)
Attention & interpretability
The Persistent Tension Hypothesis
Endogenous attention capture in large language models — how internal representational competition shapes what a model attends to, independent of prompt design. With Dr. Atif Naseer.
Attention steering
Focus Mirror
An interactive framework for steering large language model attention in real time during inference — directing and redirecting what the model focuses on without retraining. With Dr. Atif Naseer.
AI governance
Lifecycle Governance for Multi-Agent Systems
Principles and a research agenda for governing multi-agent AI systems from design through deployment, monitoring, and retirement — covering accountability, auditability, and human-oversight requirements across regulated environments.
Disaster management
Edge-Enabled Generative AI and AR/VR in Disaster Management
A theoretical framework for integrating edge AI with augmented and virtual reality to support decision-making in disaster response — relevant to critical infrastructure operators and civil contingency planners. With A.A. Khan.
Beyond the brief
The Temporal Fluid Multiverse
A unified framework for causality and observation, submitted to Royal Society Proceedings A. Outside SAIL’s applied remit — included because we publish across disciplines.
All models and datasets: huggingface.co/nshah-fbcs
·
ORCID
·
Google Scholar
Intellectual property
Patents and trademarks
Patents
2 granted US patents
Granted
Cited by Apple, Microsoft, Amazon, IBM, Morgan Stanley, Lenovo, Citrix, and Sturm Ruger & Co.
7 applied US patents
Pending
UK Trademarks
Orion — UK00004116898
Registered
eXtended Intelligence (XI) — UK00004350265
Registered
MTRX — UK00004090712
Registered
Where we publish
Our Channels
Papers and models
Preprints on Zenodo and TechRxiv. Trained models and datasets on HuggingFace. Code where it matters.
AI Business Review
Longer-form research and analysis for a business and policy audience. Independent editorial publication under AIU.ac.
SAIL Newsletter
Fortnightly on LinkedIn. Sovereign AI policy, deployment analysis, and original research for decision-makers in regulated industries.
People
The team behind the lab
Noman Shah, FBCS
Lab Director · Founder & CEO, XEROTECH LTD · Founder & President, AIU.ac
Fellow of the British Computer Society — a distinction held by the top 1% of 60,000 members. 9 patents cited by Apple, Microsoft, Amazon, and IBM. 7 published research papers. Executive education at Harvard Business School and Oxford Saïd. HBR Advisory Council. Judge for the NASA Conrad Challenge. Previously designed and led a $200M World Bank-funded technology programme across three countries. Creator of CallGPT 6X and the AIfu methodology.
Dr. Atif Naseer
Co-Founder & CTO, XEROTECH LTD
PhD Machine Learning. 60+ publications, 433+ citations on Google Scholar. Co-author on the Persistent Tension Hypothesis and Focus Mirror. Built crowd analytics systems processing millions of data points in real time. Research faculty at Umm Al-Qura University since 2012.
Industrial arm
XEROTECH LTD · Company #14474495 · ICO ZC065188
Academic arm
Artificial Intelligence University · UKPRN 10095512 · BCS-approved centre
Backed by
Google for Startups · Microsoft for Startups · Barclays Eagle Labs · AWS Activate
Commercial products
CallGPT 6X (privacy-first AI workspace) ·
VisionXI (AI accessibility) ·
PULVINIR (regulatory intelligence)
The gap is real. We are closing it.
If your organisation handles regulated data and needs AI that works inside your walls, talk to us. If you are a researcher working on small models, deployment efficiency, or sector-specific governance, we are open to collaboration.
XEROTECH LTD, Company #14474495 · Artificial Intelligence University, UKPRN 10095512