Daraja

@isaactevin @KelvRoman Thank you!

@KelvRoman @DarajaAI cool app

Two ways to build AI for African languages: 1. Translate into them 2. Be born in them One thinks in the language. The other secretly thinks in English and translates back. The difference is the whole game. - @DarajaAI, AI for Africa.

Two ways to build AI for African languages: 1. Translate into them 2. Be born in them Daraja is built entirely on the second - models that think in the language, not ones that translate toward it. Origin isn't a detail. It's the architecture.

open.substack.com/pub/reamby/p/k… - daraja.ai, AI for Africa.

"Not advocacy. Infrastructure." Rebecca Mbaya from The African Innovators Series captured exactly what we are building: Africa doesn't need AI translated into its languages - it needs AI born in them. Origin isn't a detail. It's the architecture. Full feature link below

The sharpest sentence anyone has written about what I am building: "Africa does not need AI translated into its languages. It needs AI born in them." A model that reasons in English and hands back Swahili never held the thought in Swahili. Origin isn't a detail. It's the architecture. Featured this week in The African Innovators Series (TAIS) open.substack.com/pub/reamby/p/k… - daraja.ai, AI for Africa.

Five posts this week, one argument: Sovereignty isn't a slogan. It's a choice about who owns the intelligence you depend on - your data, your language, your value. Daraja is built to keep all three at home. It's the reason we exist. - daraja.ai, AI for Africa

I spent this week arguing that "AI sovereignty" isn't a slogan. It's five concrete engineering decisions: train from scratch, don't wrap an API, run on-premise, data never leaves - learn the language, don't translate to it - own the IP, don't rent access - optimise for your users, not someone else's Sovereignty isn't a value. It's an architecture. It's a feature. - daraja.ai, AI for Africa.

The subscription fee is the cheapest part of renting foreign AI. The real cost is everything that leaves with it - the compute spend, the jobs, the IP, the tax base - all accruing somewhere else, under someone else's flag. Build locally, and the value compounds here

Every model is optimized for a distribution. Someone chose that distribution. It wasn't you. That's why "is this AI good?" is the wrong benchmark. Good at what? Measured on whose data? Rewarded for solving whose problems? An AI is never neutral. It's good at what it was built to be good at - for whoever it was built for. Sovereignty is just making sure that's you. - daraja.ai, AI for Africa

Before servers. Before compute. Before data. Whose language does the AI speak? Most systems reaching Africa answer in English and ask everyone else to adapt. Daraja builds AI that learns African languages first - not translated in afterwards. Language is the first sovereignty.

Real language sovereignty means the model learns the language itself - not a wrapper bolted on at the end. Harder to build. Also the only version that's actually ours. The Mission Continues. @DarajaAI, AI for Africa daraja.ai

Build your core on someone else's AI, and you are one policy change away from losing access to it. Daraja trains from scratch and deploys on-premise - so an institution's intelligence, and its data, remain its own. Independence that can't be revoked. daraja.ai

Everyone building on someone else's AI is one email away from losing everything. The terms change. The price 10x's. Your use case gets flagged. And you find out the thing you built your company on was never yours - you were renting permission to think. In Africa, that fragility isn't a risk. It's a certainty you're choosing to ignore. We trained from scratch so no one can send that email. daraja.ai, AI for Africa.

Using AI and owning it are not the same thing. Most African AI is access to a model built elsewhere, run elsewhere, on data sent elsewhere. daraja.ai is built the other way: trained from scratch, deployed on-premise, data that never leaves the institution.

Most African AI is a thin wrapper around someone else's model, someone else's servers, someone else's data pipeline. We went the other way. Trained from scratch. Runs on-premise. Nothing leaves the network. It's the harder build by a wide margin. Here's why we chose it anyway: you can wrap an API in a weekend, but you can't own what you wrapped. The moment the thing you depend on is someone else's, so is the sovereignty. Africa doesn't need better access to other people's AI. It needs its own. - @DarajaAI, AI for Africa. daraja.ai

One week, one case made from every angle: the smallest stroke lesion is often the most treatable - and the easiest to miss. Our model detects and outlines ischemic lesions down to 1cm, measured against the international standard. Built for the catch that hides

I spent this week explaining why we chased 1cm stroke lesions instead of the big, obvious ones. The technical part was never the hard part. The hard part was conviction - betting your effort on the small catch, the one that's easy to miss and easy to undervalue, because it's the one where intervention still changes the outcome. Down to 1cm, benchmarked to the international standard. Most models optimise for what's easy to detect. We optimised for what's worth detecting. - #MedScan Medical intelligence, built for Africa. daraja.ai

A stroke model's real value isn't catching what the radiologist misses. It's a second read on every scan - no time lost, no ego, no favour owed - so the clinician catches it themselves. #MedScan detects ischemic lesions down to 1cm, benchmarked to the international standard.

Everyone thinks a stroke model's job is to catch what the radiologist misses. It isn't. Its job is a second read on every scan - one that costs no time, no ego, no favour owed - so the radiologist catches it themselves. As sharp on the 40th scan as on the 1st. Detection down to 1cm, benchmarked to the international standard. The tireless part isn't the breakthrough. The un-offended part is. - #MedScan Medical intelligence, built for Africa. daraja.ai @UNDP, @WHO, @gatesfoundation