We build the thing,
then we say what it did.
Eullar Labs is an AI research company. We make practical tools and publish what we learn making them — the useful results and the dead ends alike.
Two rooms, one failure.
A teacher with forty students cannot diagnose forty different misconceptions in one period, so the class moves on and the gap compounds. A hiring team with four hundred applications cannot write four hundred honest replies, so it sends a template and four hundred people learn nothing.
Both are the same shape of failure: attention that does not scale, in a place where the individual case is the only thing that matters. That is a narrow, tractable problem for machine learning — and a badly served one, because the systems built for it are usually optimised for engagement rather than for the outcome that justified building them.
We started Eullar to work on the version of these problems that shows up in a real classroom in Accra and a real application queue — not the version that fits neatly in a benchmark.
That commitment sets the shape of the company. The research team is the product team. Evaluations run against deployments. And every system we ship has a documented limit on what it is allowed to decide on its own.
Six commitments we can be held to.
Written to be falsifiable. If we break one, it should be obvious from the outside.
Research earns its keep by shipping
We do not maintain a wall between the lab and the product. A result that cannot survive real users, real latency and real edge cases is an unfinished result.
The human keeps the decision
Our systems explain, propose and draft. Teachers keep authority over a learning plan; hiring teams keep authority over a hiring outcome. We build the seams that make override easy.
Specificity over reassurance
Vague output is a way of hiding uncertainty. We would rather a system name a narrow gap it can defend than produce a fluent answer to a question it did not understand.
Measure the thing, not the proxy
Satisfaction scores, engagement time and completion rates are proxies. We instrument for retained understanding and acted-on feedback, even when they are slower and less flattering.
Publish the failures
Our working notes include the methods that did not hold up. A lab that only publishes wins is a marketing department with a LaTeX template.
Build where it is needed
We are based in Accra and build first for classrooms and job markets that most AI products treat as an afterthought. Constraint is a design input, not an excuse.
Model-agnostic by design.
We do not train foundation models. We build the representation, evaluation and control layer above them, and we keep the inference layer swappable so a product decision is never hostage to one provider's roadmap.
Asked often, answered plainly.
No. We build the layer above them — representation, evaluation and control — and we stay deliberately model-agnostic so that a product decision is never hostage to a single provider's roadmap.
Customer content is processed for the customer's own workload and is not used to train shared models. Schools and employers keep their data boundaries; deployment options include region-pinned and self-hosted inference.
No, and it is architecturally prevented from doing so. Reevue receives an outcome and explains it. Selection stays with the hiring team.
It is a planning and diagnostics instrument for teachers. Every route it proposes is inspectable and editable, and the teacher's override is a first-class action rather than an escape hatch.
Yes — particularly with education faculties and labour-market researchers who want to run studies on deployed systems rather than on benchmarks.
Accra, and wherever the deployment is.
Building from West Africa is a design constraint before it is an origin story: intermittent connectivity, large classes, price sensitivity, and multilingual classrooms are the default case here. Systems that hold up under those conditions tend to hold up everywhere else.
- founded
- 2024
- base
- Accra, GH
- latitude
- 5.6037° N
- longitude
- 0.1870° W
- products
- Syllabi · Reevue
- team
- research = product