Research — § 00 / Three active · One in development

Can language be trusted?

Three active research programmes and a fourth in development — each led by named researchers, each producing evidence that can be inspected, reproduced, or challenged by the people reading the output under regulatory oversight.

Status · 2026 Q2
Programmes03
Named leads05
In development01
§ 01 — ProgrammeKTP · Sheffield Hallam

Multilingual data as evidence.

Two Knowledge Transfer Partnerships with Sheffield Hallam University, recognised among the top 50 KTPs in UK history. The first produced GAI Translate. The second develops agentic AI methods for multilingual dataset labelling in sensitive, regulated domains.

LeadAlex Shenfield · Chris Roast (SHU)
Specimen setMultilingual labelling corpus
Product surfaceGAI Translate
Active work
W / 01 KTP 01 — originating the GAI Translate product line Shipped · 2023
W / 02 KTP 02 — agentic AI for multilingual dataset labelling Active · 2026
W / 03 Recognised among top 50 KTPs in UK history — Innovate UK Award · 2025
§ 02 — ProgrammeSmall Language Models

Domain beats scale.

Domain-specialised small and medium language models, trained on a human-verified data lake and deployed on secure Microsoft Azure private cloud. Our thesis: for regulated sectors, a small model trained on the right corpus outperforms a general-purpose LLM under audit.

LeadIsham Iqbal · Shashwat Upadyay
Specimen setHuman-verified domain data lake
Product surfaceGAI Translate
Active work
W / 01 World-first mining-specific SLM Shipped · 2025
W / 02 SLM training for an enterprise client — regulated finance proof-of-concept PoC · 2025
W / 03 Azure private-cloud deployment with tenant isolation Infra · Live
§ 03 — ProgrammeSynthetic Audiences

Stress-test language before it ships.

Multi-persona agent panels that read translated or generated language the way a Japanese pharma regulator, a German finance counsel, or a Gulf compliance officer would — surfacing precision, tone, and compliance failures before the output reaches a real reader.

LeadDr. Francimaria Santos · Shashwat & team
Specimen setMulti-persona evaluation panels
Product surfaceEvaluation layer (beta)
Active work
W / 01 Multi-persona synthetic audiences for precision evaluation Method · Active
W / 02 Cross-jurisdictional compliance-reader agents Build · Q2
W / 03 Productising synthetic audiences as a standalone evaluation layer Market · 2026

A fourth programme is being formalised — Agentic AI research with our internal AI team.

Scope, named leads, and first outputs will be published later in 2026. Write to [email protected] to be notified when the programme goes live.

P / 04
Agentic AI research for regulated-industry workflows

Joint work with Guildhawk’s AI team. Formal programme launch, named leads, and first evaluation set in a future update.

Programme Coming soon