K
Kato Steven Mubiru
Crane AI Labs / Cohere Labs Community
Co-founder of Crane AI Labs and co-lead of the Regional Africa community at the Cohere Labs Open Science Community.
How a Luganda word-review task became a startup, a Gold Award, and a lesson about what really keeps gates closed for African builders, and what opens them.
More than a year ago, an email arrived asking if I could review a list of Luganda words. Simple task: go through a spreadsheet, flag what was wrong, make sure the data was clean. That email was from Alejandro Salamanca at Cohere Labs. I said yes, not knowing it would change the direction of my life.
I am Kato Steven Mubiru, co-founder of Crane AI Labs and, alongside Bakunga Bronson, a lead for the Regional Africa community at the Cohere Labs Open Science Community. This is the story of how showing up in a Discord server became the foundation for everything we’ve built, and what it taught me about the gates that stay closed to African builders, and what actually opens them.
It was not a big break. Nobody handed us a grant or a stage. It was a spreadsheet. But we treated that spreadsheet like it mattered, every word checked, every flag explained, because it was our language on the line, the words my grandmother speaks. That work fed into the Aya training pipeline, and later into models that now serve African languages.
Looking back, that is the whole point. The gate didn’t open because of who we were. It opened because a small task was done to standard.
Here is something I believe now that I did not understand then: the gates that stay closed to African builders are rarely closed because of talent. Talent is everywhere here; I see it every week in our community channel. What keeps gates closed are small things. The email that never gets a reply. The dataset released without documentation. The claim that doesn’t survive a careful read. The deadline that slips without a word. None of these are about ability. All of them are about standards, and standards can be learned.
That is what Cohere Labs actually gave us. Not just access, but apprenticeship. When we ran evaluations, mentors pushed us to report what the data said, not what we hoped it said. When we wrote up our work, we learned that crediting every contributor is not politeness; it is rigor. When we shipped models, we learned the README matters as much as the weights. Quality compounds. Every carefully finished small thing became the reason someone trusted us with a bigger one.
Takeaway: Quality compounds. Every carefully finished small thing becomes the reason someone trusts you with a bigger one.
I kept showing up. Then contributing more. Then, slowly, leading. Bronson and I became leads for the Regional Africa community, responsible for a space where African researchers and builders could gather, learn, and be seen.
We taught ourselves the art of cold outreach and invited researchers the community would never otherwise meet: David Vilar from Google DeepMind on TranslateGemma, and Pedro Ortiz Suarez and Laurie Burchell from the Common Crawl Foundation. And we learned quickly that our members didn’t just want talks. They wanted to be seen. Dedicated mentorship is why most people stay in this community; it tells them their work matters. That insight shaped how we led.
In 2025, Bronson and I co-led Afri-Aya during Expedition Aya: a culturally grounded vision-language benchmark covering 13 African languages, seven appearing in such a benchmark for the first time, built with annotators across the continent who understood the languages and cultures from the inside.
Afri-Aya won the Gold Award at Expedition Aya 2025. The prize mattered less than the proof: a team working from Kampala, holding itself to the community’s standards, could produce research that stood alongside work from anywhere in the world.
Everything we learned flowed into Crane AI Labs: the rigor, the storytelling, the network. When we wanted to prove that small models could do real work on African phones, we returned to the community for Expedition Tiny Aya.
In two weeks, our team built Tiny Facade, an open-source Android service that lets any app run a multilingual AI model on a phone, offline. A 3.35B model with no tool-calling training outperformed a model more than twice its size on Luganda. Along the way we found a bug in a globally used inference tool that was quietly degrading performance for every user, and fixed it, releasing corrected models for the whole community. Julia Kreutzer from Cohere mentored us through the research and pushed us to ship something rigorous, not just fast.
Our work at Crane has since been featured by Google DeepMind in their official Gemma showcase. But it all started here, in a Discord server, with a word-review task done carefully.
So this is my call to African builders: own your space. Join communities like this one. But do not just attend. Hold the standard while you build. Review the spreadsheet carefully. Document the dataset. Answer the email. Show up when you said you would. Make honest claims and let the work carry them.
The world does not open its gates because we ask. It opens them when the work is undeniable, and undeniable is built from small things done well, over and over, until nobody can look away.
Bronson and I have carried this community, but it has grown beyond what two people can hold. Over the coming months we are mentoring the next generation of Regional Africa leaders: monthly sessions every third Tuesday, and a deliberate search for members who show up, who care, and who hold the standard. We are stepping into the role our mentors played for us: opening doors, making people seen, and trusting them to lead.
Because that is how this community works. Someone believes in you before you believe in yourself. Then one day, you get to be that person for someone else.
A huge thank you to Sara Hooker, Marzieh Fadaee, Madeline, Brittawnya, Alejandro Salamanca, Julia Kreutzer, and the entire Cohere Labs team for building this community and continuously raising its bar. And to every member of Regional Africa: this is your story too.