A small product studio in Abu Dhabi, building on Western and Chinese frontier models for enterprise and developer markets.
Film, 24 seconds: the camera pulls back from a light in Abu Dhabi to orbit, where six frontier models circle the planet and requests are routed between them.
In stealth. Names and demos follow. For now, the shape of the work and what we think it's worth building.
A model-routing runtime for teams that don't want to pick a provider.
Intercepts your chat and tool-use calls, picks a model per request based on latency, cost, competence, and residency. Works across Western and Chinese frontier models out of the box.
A developer SDK for building agents that behave themselves.
Typed tools, retry and fallback policies, structured evals baked into the call path. Written in TypeScript, works in Node, edge runtimes, and browsers. In pilot with a handful of design partners.
An enterprise workspace that turns internal documents and systems into working AI agents.
Built for government and regulated enterprises — policy documents, case files, operational handbooks — and the consulting teams that deliver them. Secure by default, deployable in-region, designed to slot into existing workflows rather than replace them.
We route across Western and Chinese model providers on a per-task basis. Latency, cost, and the specific competence of the underlying model decide the call — not a procurement preference. The orchestration layer is the product.
Abu Dhabi is our home, and a good vantage point. The Gulf sits between the three regions we work across, and ADGM gives us serious legal footing. The result is a studio that feels local where it needs to be, and global where it can be.
Small teams ship faster than big ones. We keep headcount low on purpose, and treat research as something that happens in the same file as the code. The distance between an idea and its deployed version is the thing we optimize.
Not a blog. Short margin notes, filed under the problem they belong to.
A paper we keep returning to: the marginal model is usually not the best model. Picking the cheapest competent model per call is a larger win than most benchmarks admit.
The interesting part of data-residency is not where the weights live. It is where the prompt lives, where the cache lives, and who is allowed to see the intermediate reasoning trace.
Tokenization quietly punishes non-Latin scripts. A seven-word Arabic sentence can cost twice what its English translation costs to generate. A lot of our internal work is this kind of invisible tax.
The Chinese open-weight ecosystem is closer to the frontier than most Western teams have internalized. Several of the labs there are in our production rotation.
The Gulf is the shortest flight between the three regions we care about. ADGM gives us common-law footing, and the city is quietly becoming one of the more interesting places to build AI infrastructure that takes regulation seriously.
For research partnerships, product collaborations, or careers.
Al Maryah Island · ADGM · Abu Dhabi, UAE