African AI research lab
We advance AI research in Africa. We study how models learn from human judgment, we build and test AI tools, and we run the platforms that produce the data behind them.
Why we exist
To advance AI research in Africa by doing the research here, building the tools it requires, and creating the conditions for African researchers to work at the frontier.
An Africa that shapes how artificial intelligence is built, not only how it is used, with the research that makes that possible published from here.
Research
Model capability now depends less on architecture than on the quality of the signal a model learns from. Our work sits on that signal.
Reward modelling and preference elicitation, and what makes a ranking signal reliable rather than merely consistent.
Benchmarks that measure what they claim to, and methods for detecting when a model has learned the test rather than the task.
Speech and text for languages with little written corpus. Code-switching, dialect variation, and transfer between related languages.
Interactive settings where a model attempts long-horizon work and a written rubric scores the attempt.
Red-teaming and safety evaluation in the contexts and languages mainstream testing does not reach.
What happens when these systems meet real institutions, studied with the people who work in them.
Practice
Research that never leaves the paper is only half the work. We develop applied AI tools, and we test other people's.
Retrieval systems, agent workflows, language tooling, and the evaluation harnesses that keep them measurable as they change.
Capability probing, adversarial testing and regression suites, assessed by specialists in the domain the tool claims to serve.
The pipelines and measurement that turn expert judgment into data a research team can train on. This produced Peertrail.
Approach
A lab that depends on someone else's data pipeline can only study what that pipeline produces. We build our own.
When a study needs a different kind of signal, we build the instrument to capture it rather than working around what exists.
Every dataset ships with the numbers behind it: assessor agreement, accuracy against known answers, and where coverage is thin. Work we cannot measure, we do not claim.
Clinical questions answered by clinicians, language questions by native speakers. Depth in the subject beats familiarity with machine learning.
Platforms
Some research questions need infrastructure that does not exist yet. Where that happens we build it.
Peertrail is where specialists do the assessment work our research depends on. Members qualify for a project, complete paid tasks, and have their work reviewed independently by others in the network. Nothing counts as data until independent assessors agree on it, and that agreement is measured rather than assumed.
Peertrail is open to members in Nigeria while the first cohorts run. Further markets follow.
Contact
Commissioning research, proposing a collaboration, or asking whether your background fits a project. A person reads everything that arrives here.