We built a Tigrinya AI model.
A summarisation model trained on 2.8 million words of Tigrinya. Benchmarked against GPT-3.5 and GPT-4.1. Open source.
Translational Justice
Language is failing at what it should do. Communicate.
In asylum hearings, bad interpretation becomes legal fact. AI reads some languages worse and charges others more. Autocorrect treats dialect like a mistake.
That cost has a shape.
CF-01 / Asylum
Utterance vs. record
"I cannot say this here" rendered as I do not know .
Tribunal finding: claim not credible.
Active investigation
CF-02 / Language Technology
Google Translate, autocorrect and Singlish
The words can be right while the human connection disappears.
Fieldwork active
CF-03 / AI systems
Low-resource language bias
Automation does not flatten the hierarchy. It encodes it.
Research ongoing
What exists now
Research. Policy. Teaching.
The Policy Briefing.
Two fixes for asylum interpretation: qualified interpreters, and original-language records that can actually be checked.
The Classics Briefing.
A proposal for teaching classical texts without pretending the translation is the original.
What the translation killed.
Six words from Pliny Letters 9.33. Word by word. What the English cannot carry.
TEDx Talk - The foundation
The Words I Had to Leave Behind
The investigation starts here. A talk about what meaning loses when systems try to carry it.
Watch on YouTube
External blogs now live
Two new pieces are live.
Policy briefing / June 2026
Download PDF
Interpretation Failure in the UK Asylum System
Two reforms: mandatory interpreter accreditation, and original-language testimony kept beside the English record.
Case studies
Record
Asylum. Singlish. AI.
Three places where systems decide what meaning survives: hearings, keyboards, and AI.
Data can be ignored. Experience cannot.
A WebXR simulation inside an asylum hearing. You speak. The translation drifts. The record hardens without you.
Enter simulation FieldworkLawyers. Interpreters. Standards bodies.
Direct conversations with lawyers, interpreters, standards bodies, scholars, and refugees. Names removed. Patterns left in.
Read field notes WritingEssays, papers, and public-facing work.
Procurement chains. Singlish particles. AI language hierarchies. Same problem, different rooms.
Browse writing