A wrong answer in most AI is an inconvenience. You laugh at the chatbot, refresh, move on. A wrong answer in a cancer screen is a funeral. I had to sit with that early. When you build a model that tells a clinician "nothing here," you're not tuning a metric anymore. You're deciding, in some small way, who gets caught in time and who gets sent home. So I stopped chasing accuracy. Accuracy is a vanity number. It rewards a model for being right about the easy cases - the healthy scans, the obvious negatives. A model can be 95% accurate and still miss the one tumour in the room, because the one tumour was rare and the metric didn't care. I build against the opposite mistake. The miss. The false "you're fine" handed to someone who isn't. In medicine that's the only error that kills - so that's the one I optimize against, even when it makes the headline number look worse. It means my models are sometimes more cautious than they strictly need to be. They flag things for a human to check that turn out to be nothing. I'll take that trade every time. A second look costs a clinician a minute. A missed cancer costs a life. People think the hard part of medical AI is making it smart. The hard part is making it humble - teaching a machine to fear the mistake that matters more than it loves the score that doesn't. That's the standard. Not because it's clever. Because someone's mother is on the other side of the result.

We talk about food sovereignty. Energy sovereignty. We need to start talking about cognitive sovereignty. Every nation is racing to own its AI. But ownership isn't just data centres and chips. It's language. A model that can't think in your language can't serve your people - it can only translate them, flatten them, approximate them. Right now a handful of companies on two continents are deciding which of the world's 7,000 languages are worth teaching a machine. Most African languages didn't make the list. That isn't a technical accident. It's a decision about who the future is for. Africa cannot outsource that decision. We need to build what only we can build. The future will speak many languages. Ours had better be among them. daraja.ai #AIforAfrica

"Niko poa kabisa" daraja.ai #BabelVoice #AIForAfrica

In Builders Space: Building Products in Africa's Tech Ecosystem Episode 2 we had a chat with @KelvRoman and @opere_brandon on Developer to Builder: Moving Beyond Tutorials Into Real Products here are the takeaways linkedin.com/pulse/developer…

23 languages. That was the text layer. But 60% of Africa's internet users prefer voice over text. And millions more never learned to read in any language. Text-only AI excludes the people who need it most. So we taught the bridge to speak. Voice is coming to @DarajaAI. Not as an add-on. As a core layer. Translation, conversation, medical screening - all delivered in the language you think in, spoken in a voice you understand. daraja.ai #AIforAfrica

2 AM. A GPU somewhere is learning how to pronounce Yoruba. 3 AM. Another one is learning Pidgin. 4 AM. I'm reading training logs the way a new parent watches a baby monitor. There's no shortcut to teaching a machine to speak a language. You feed it thousands of hours. You clean every sample. You train, fail, retrain, fail differently, retrain again. But when it finally says "Habari yako" and it sounds right - that silence after is the best sound in the world. Voice is coming. daraja.ai #AIforAfrica

User feedback drives everything we build. When our users speak, we listen - and we ship. Thank you. daraja.ai #AIforAfrica

"Bonzour Babel" That was the first sentence ever spoken to an AI in Mauritian Creole. I typed it into a model nobody knew existed, for a language nobody had built for. Babel replied in Creole. Naturally. Like a person would. That was one language. Today it's 23. Twelve African. Five European. Six Asian. Any language to any language. Interconnecting Africans and the world. Every revolution starts with one sentence nobody thought was possible. daraja.ai #AIforAfrica

I love many aspects of being an engineer but one thing that stands out is user feedback. Nothing really comes close to it. You get to understand where the user's mind is at, how they interact with your system, and most importantly - where you got it wrong. That data is invaluable. No amount of internal testing replaces it. This morning a user flagged something. Within minutes it was addressed and deployed. Not because I'm fast. Because when someone takes the time to tell you what's broken, the least you can do is take it seriously. Every piece of feedback makes the system better. Every bug report is a contribution. Every 'this doesn't feel right' is someone who believes in what you're building enough to help you build it better. We are listening. We are improving. Every single day. daraja.ai #AIforAfrica

As Africans, we have always passed down knowledge and wisdom through oral tradition - from the scholars of Timbuktu to the storytellers of today. So when we built Babel, we refused to give our models cold, forgettable codenames. Each one carries the name of a son or daughter of the motherland - a poet, a playwright, a linguist, a keeper of stories - who loved their language so fiercely that history could not forget them. Babel-Virahsawmy honours Dev Virahsawmy, who spent his whole life insisting that Mauritian Creole was a language and not a lesser tongue - translating Shakespeare, the Bible, the Bhagavad Gita and the Quran into Creole so his people could meet the world in their own voice. Mariama Bâ, who gave Senegalese women a literature of their own. Chinua Achebe, who made the world read Africa on Africa's terms. Hadraawi, the "Master of Speech," who would not stop writing the poems that put him in prison. They were told their languages were too small to carry literature, science, scripture. They refused to believe it. We carry that same refusal into the age of AI. Every time someone speaks to Babel in their mother tongue and is understood, the work these men and women began continues. We didn't borrow their names for prestige. We took them so that a child using Babel might one day ask, "Who was Virahsawmy?" - and go and find out. Their legacy lives on. Now it answers back, in the languages of the continent. daraja.ai/explore @DarajaAI #AIforAfrica

This is how tafsiri works with @DarajaAI Project: github.com/ianktoo/tafsiri

Came across @DarajaAI and their African language models while building Crea (an emergency response app). That stopped me. Language is a barrier in emergencies, a wrong translation can be worse than none. So I built tafsiri. An eval pipeline that runs on @DarajaAI API as translation engine. Right now, it translates into Swahili, Yoruba, Amharic (rides on Daraja). Evaluates each output confidence (Daraja's own confidence), back-translation (asks Daraja the reverse translation) and LLM-as-judge (any frontier large language model). Not just a translation. A score you can trust. It's < 5 days old. Open source. No stats yet. Try it out! github.com/ianktoo/tafsiri