Concept stage. Cenab is not built yet — this page describes what we intend to build, and we are looking for pilot partners to design it with.
Your agent speaks 20 languages. You've tested one.
Cenab will measure how your AI agent actually behaves in every language you ship to, score each one against your English baseline, and produce the evidence trail auditors now ask for.
A model card lists languages. It does not promise quality.
A 2025 study deliberately searched for questions a model answers correctly in English and gets wrong in another language. Across sixteen languages it found thousands of them, and on those cases most models lost more than fifty percentage points. The failures are there, they are findable, and almost nobody goes looking.
So teams ship an agent into twenty markets on the strength of English testing and a vendor's language list. The first time anyone learns the agent gives wrong prices in Turkish is when a customer acts on one.
Agent accuracy by language
IllustrativeAn example of the report Cenab would produce. Each language scored against the same agent's English baseline, acceptance threshold 85.
What Cenab will do
Tests in every language you ship
You give Cenab your agent endpoint and your own policies. It builds and runs test suites per language: factual accuracy against your source of truth, refusal behaviour, tone, terminology and prompt injection.
Scores against your English baseline
Every language is measured against how the same agent behaves in English on the same cases. You see the gap, not an abstract score you cannot act on.
Produces the evidence
Each run keeps inputs, outputs and verdicts per language, exportable as a report. Wire it into CI and a release stops when a language drops below your threshold.
How it will work
Four stages. You keep your keys and your data.
The evaluator is the hard part
Grading an Azerbaijani reply needs a model that reads Azerbaijani well. The plan is to use Claude as the default evaluator, with a different model pinnable per language.
Why this is urgent now
The EU AI Act has been fully applicable since 2 August 2026. Language systems used in hiring, legal work, healthcare and public services can fall into the high-risk class, where penalties reach €35 million or 7% of worldwide turnover.
Your data stays yours
Access keys stay in your accounts. Test runs can execute against a staging endpoint, and what is retained is agreed in writing before the first run.
Where we start
Turkic, Slavic, South Caucasus and Middle Eastern languages first — the ones we have built agents in ourselves and where the measurement gap is widest.
Looking for the first ten pilot partners.
Nothing is built yet. We are looking for teams already running an agent in more than one language, who want to know what it is doing in the languages they cannot read, and who will help shape the first version.
What a pilot would include
- A measured report across your live languages, once the first version runs
- The failing cases, with the exact inputs that produced them
- Direct access to the founder through the pilot
- No cost during the pilot
What we ask
- A staging endpoint we can call
- Your policies and the facts the agent must not get wrong
- An hour of your time at the start and at the end
- Honest feedback, including that it is not useful
Company
CENAB LLC was registered in New Mexico, USA in June 2025. The team works from Baku, Azerbaijan, and has been building production AI agents in low-resource languages since then.
Before this, and still today, the team builds and runs WhatsApp AI agents for businesses in Azerbaijan — answering customers, qualifying enquiries and booking appointments in Azerbaijani, on the same Cloudflare Workers and Claude stack. That work is live, and it is where the language problem behind Cenab came from.
- Legal name
- CENAB LLC
- Registration
- New Mexico, USA — Entity ID 0008032490
- Founded
- 18 June 2025
- Registered address
- 1209 Mountain Road Pl NE, Ste N, Albuquerque, NM 87110, USA
- Operations
- Baku, Azerbaijan
- Founder
- Sayyad Yusifov
- Contact
- founder@cenab.online
- Team's stack
- Cloudflare Workers, Claude API