Case study
← All workGraysen
An AI contract analyser scoped to named UAE legal instruments, working in Arabic and English.
- Client
- Graysen
- Year
- 2026
- Discipline
- Applied AI · Legal
- Live
- www.graysen.ai
Context
Most people sign contracts they have not fully understood, and in the UAE the gap is wider than usual: the governing instrument depends on whether the agreement is employment, tenancy or services, and on whether it sits under federal law or inside a financial free zone with its own common-law framework. The product had to hand a non-lawyer the specific clauses that will cost them, with the reasoning attached.
Constraints
Jurisdiction was the whole problem. A model reasoning from general legal training produces answers that are fluent, confident and wrong in the Emirates. The product also had to work in Arabic and English where neither is a translation of the other — an Arabic contract is the binding document, not a copy — and it had to stay useful to someone with no legal training while staying on the correct side of a line that, once crossed, is unlicensed legal practice.
Architecture
Six services in production: the web application, a headless CMS on its own service so legal content can be revised without a deploy, a backend API, two PostgreSQL instances and Redis. The separation of CMS from application is the load-bearing one — the frameworks the analyser reasons against change when the law changes, on a schedule nobody controls, and that must never require an engineer.
01
Scope the model to named instruments rather than to law in general.
A general legal reasoner cannot be checked. Binding the analysis to nine specific named frameworks — and publishing that list on the marketing page — makes the product's competence inspectable: a reader can see the ground it stands on, and therefore judge when they are standing off it.
What it cost
The product is useless outside the UAE without repeating the work for another jurisdiction. Deliberately narrow.
02
Return discrete flagged clauses, not a written opinion.
A wall of prose reads as advice and invites a reader to trust it wholesale. A list of specific flagged clauses, each with a plain explanation, is scannable by someone who has never read a contract closely — and structurally makes a narrower claim.
What it cost
Findings that span several clauses are harder to express, and the format cannot easily say 'this contract is fine'.
03
Carry the analysis-not-advice disclaimer in the hero, not the footer.
The regulatory line is the product's actual boundary. Putting it where every visitor reads it is both the honest choice and the defensible one.
What it cost
A caveat in the first screen costs conversion that a footer disclaimer would not.
- Services in production
- 6
- read from deployment
- Datastores
- PostgreSQL ×2 · Redis
- read from deployment
Live at graysen.ai on a custom domain, bilingual throughout, naming nine UAE instruments as the ground its analysis stands on.
Stack
TypeScript · Next.js · PostgreSQL · Redis · Headless CMS · Google OAuth