Summary
- Scottish company Vidai UK says Stirling-Minimo PII can identify personal information without sending text to an external detection service.
- The company describes a 10-million-parameter model of about 43MB that can run on ordinary CPU hardware.
- Non-commercial use is available under a Creative Commons licence; commercial use requires a separate agreement.
Full Article
Scottish AI company Vidai UK has developed a compact model designed to identify sensitive personal information before it is sent to external AI providers. The company says Stirling-Minimo PII runs within an organisation’s own infrastructure, allowing data checks without transmitting the text to a separate cloud-based detection service.
Daily Business reported on the model on 28 October, describing its use in identifying names, addresses and financial account numbers for redaction. Vidai’s own model documentation on Hugging Face describes a system with 10 million parameters and a file size of about 43MB, designed to run on CPU hardware without requiring an accelerator.
Nagu Gopalakrishnan, co-founder and CPTO of Vidai UK, said, as reported by Daily Business: “Sending sensitive information to an external cloud service just to check if it contains private data completely defeats the purpose of privacy protection.”
He added: “We built Stirling-Minimo in Scotland to give organisations absolute authority over their own data footprint. Scotland is proving to be an exceptional hub for deep-tech innovation, proving that world-class, privacy-first AI can be built right here rather than imported from Silicon Valley.”
The company says the model identifies 77 types of entity, including contact details, financial information and credentials. Its v0.2 release adds processing for structured material such as XML, SQL and logs alongside ordinary prose.
Vidai distinguishes the published model from its wider Control Plane platform. The downloadable model supplies detection outputs, while functions including policy enforcement, masking and audit records are handled within the wider system.
The company also lists limitations: the model is English-only, some uncommon names may be missed or fragmented, and longer documents need to be split into smaller sections. Its documentation says organisations remain responsible for compliance and advises against relying on the model alone where a missed detection could cause serious harm.
In Case You Missed It:
Glasgow 2026 closes on privately funded model as Chamber points to investment dividend
Scotland sends 60% of its venture capital into deep tech – now it needs a model built for Scottish scale
Plans submitted for 1,000MW data centre at former Chapelcross nuclear site
Hymans Robertson partners with Heka to strengthen pension data tracing
The model predates this week’s coverage. Vidai’s public repository histories record the original upload on 30 September and the v0.2 release on 13 October.
According to the company’s model cards, research, evaluation, personal and non-commercial use are permitted under the CC BY-NC 4.0 licence. Commercial use requires a separate licence, and the documentation says the model’s weights and output format may change before version 1.0.
