Small models, big continent
The cloud is not coming to every village, clinic and farm. Why everything we build for Africa is designed for the edge: offline, on-device and on the power people actually have.
Picture a farmer in northern Ghana walking her rows of maize late in the afternoon. A few plants have pale, spotted leaves, and she wants to know why before the problem spreads. She photographs a leaf and asks an app on her phone. One bar of signal, then none.
If the intelligence behind that app lives only in a distant data center, her question never leaves the field. She can try again tomorrow, walk to the nearest town or guess.
Edge-first means the answer happens on the phone in her hand. The model runs on the device and works without a connection. When the signal returns, it syncs. The cloud still helps, but nothing essential waits for it.
That is why everything we build for Africa at SISU AI starts at the edge, and why we believe every AI solution for the continent should. It is tempting to see edge-first as a fallback for places the cloud has not reached yet. We think that gets it backward. Edge-first is the better architecture, and what gets built under hard constraints tends to travel.
What the cloud assumes
Cloud-first AI rests on a few quiet assumptions: a steady connection, cheap data, reliable power, a capable phone and a data center close enough that the round trip feels instant. Across much of Africa, none of them can be taken for granted.
Start with connectivity. Many people live within reach of a mobile broadband signal but don’t use mobile internet, often because of the cost of data and devices. When data is expensive, every trip to a server has a price, and the user pays it.
Then power. Roughly 600 million people in Africa live without electricity. Where power is scarce, a phone’s battery is something to ration, and software that drains it spends something people cannot easily replace.
Then the phone itself. Many of the devices in people’s hands are affordable Android phones with limited memory. A model that needs a flagship phone will not run on the devices where it could help most.
Finally, distance. When the nearest data center is far away, every request spends time traveling there and back. In a chat window, that delay is an annoyance. In a clinic, or on a machine that has to react in real time, it can decide whether the tool is useful at all.
The people the cloud leaves out don’t need less from AI. They need AI that comes to them.
Seven principles we build by
Edge-first is a set of engineering choices, made early and kept. These are the seven we hold ourselves to.
- Offline by default, sync when it helps. The core of a product should work with no connection at all. When a signal appears, the app can sync in the background or hand off a heavier request. A connection should make the product better, and losing one should never break it.
- Small, specialized models. A model built for one job can be far smaller than one built for every job. Techniques such as quantization, which stores a model’s numbers at lower precision, and distillation, which trains a small model to learn from a larger one, help make useful models practical on an affordable phone. We look for the smallest model that does the job well.
- A capability cascade. Work runs as close to the person as possible: on the phone first, then on a nearby edge server, perhaps in a clinic or a school, and in the cloud only for the heaviest jobs. If a higher tier is out of reach, the product steps down gracefully instead of failing.
- Every channel counts. Not everyone has a smartphone, and not everyone wants to type. USSD (the menus you reach by dialing a short code) and SMS work on basic phones and don’t need mobile data. Voice serves people who prefer speaking to typing.
- Privacy by architecture. Data that never leaves the device can’t leak in transit. When a health question, a voice note or a photo of someone’s farm is processed on the phone, it can stay with the person it belongs to. Privacy then comes from how the system is built, before any policy is written.
- Energy-aware design. Battery and power budgets are product requirements, set alongside accuracy and speed. That means keeping models small enough to finish quickly and saving heavy syncing for when a phone is charging or on a cheaper connection.
- Measure what matters. We judge our work by how fast it runs on an affordable phone, how accurate it is in local languages and what each answer costs. Models read text in small pieces called tokens, and research has shown that tokenizers split many non-English languages into far more of them than English. That “token tax” can make the same answer slower and more expensive. A leaderboard score earned in English on data center hardware says little about any of this.
Where it shows up in our work
These principles already shape work across SISU AI’s companies, at different stages. The clearest examples are about reach. Dinera, which applies AI to food safety and food security from farm to fork, has an app built to work online and offline, plus access via USSD. Vaxus, which brings AI to carbon accounting and climate awareness, supports USSD and SMS so people can reach it online or offline.
Some of the work sits in the foundations. Altera, our education and mentorship company, lists “edge aligned” among its principles and fine-tunes language models for personalized learning. Alkebulan, our Pan-African large language model, is multilingual by design, and its v0.0.17 “Accra” release added foundational support for edge-friendly deployment.
Some of it is about trust. ROS, the agentic web browser from Rosanne AI for Mac and iPhone, keeps your history encrypted on the device, with no account needed. That is privacy by architecture: your history stays with you.
And some of it is physics. Pibotix, our physical AI and robotics company, sums up its work in three words: “Perceive. Reason. Act.” A robot has to do all three locally, because real-time control can’t wait for a round trip to the cloud. Robotics may be the clearest case for edge-first design there is.
Why the edge travels
For us, edge-first starts with the conditions above, but its advantages are not regional. An answer computed on the device has no network round trip to wait for. Sensitive data can stay where it was created. A product that works offline keeps working when networks fail, whether in a village or in a city during an outage.
An answer should never depend on how many bars you have.
The economics travel too. Work done on a phone someone already owns adds nothing to a server bill, so each answer costs less to deliver. Less energy goes into moving data across networks and in and out of data centers.
Ideas born under constraint have traveled before. M-Pesa launched in Kenya in 2007 and helped make mobile money mainstream across the continent. It also changed how many people far beyond Kenya think about payments. Such ideas travel because constraints that are extreme in one place exist, in milder forms, almost everywhere.
Scale points the same way. Africa is projected to reach roughly 2.5 billion people by 2050, about a quarter of humanity, so building for this continent means building for a large share of the world. SISU AI exists so emerging markets can be architects of the shift AI is bringing, and the edge is one of the clearest places to start.
Start with the hardest phone
If you are building AI for Africa, or for anywhere, start with the hardest conditions. Assume an affordable phone, a weak or missing signal, a tight battery and a person who would rather speak than type. Then add capability as conditions improve. It is usually easier to scale a product up to a fast connection than to squeeze a cloud product down onto a phone that is often offline.
Open tools and models make this easier, because anyone can adapt them to local languages and shrink them to fit local hardware. Each of our six companies has a public, MIT-licensed repository on GitHub. Start at our developers page, take what helps and tell us what breaks.
Back in the field
Return to the farmer in northern Ghana. In the edge-first version of her afternoon, the phone reads the photo and answers right there among the maize, in her language, with no signal at all. When she reaches town and the signal returns, it syncs. She never waits for the cloud to find her.
Our tagline is “Intelligence without borders.” For her, the border that matters is the place where the signal fades, and help that stops there is help she can’t count on. Build for the far side of that line, and you have built for everyone. Small models, big continent.