AI engineering · consulting · transformation
The models are being built elsewhere, at a cost no single African institution can justify. The value is not in the model. It is in the system built on top of it — the one that holds your data, your rules and your processes. We build that system, and we build it so that no single AI vendor can hold you hostage.
Africa’s share of the world’s data-centre capacity, with about 18% of the world’s people.
Africa Data Centres Association, 2026
The cost of one training run for a top-tier model today. That cost has grown about 3.5× a year since 2020.
Epoch AI — frontier training-cost estimates, 2026
Value AI could add to Africa’s economy by 2030 — almost all of it from applying models, not building them.
GSMA — AI for Africa
The situation, in plain terms
One frontier training run costs more than most African banks earn in a year, and the cost of a run has grown about 3.5× a year since 2020. Africa holds 0.6% of global compute; the United States alone holds about 45%. Even if every announced African data centre is built, the continent’s share stays flat, because building elsewhere is accelerating faster.
Africa Data Centres Association, 2026 · Epoch AI
Almost every AI system running in Africa today sends its most sensitive data to one foreign provider, in one foreign region, under one foreign policy. If that provider changes its price, changes its terms, is blocked in your region, or simply fails, the institution that built on it has no second move. A bank, a ministry or a telecom operator cannot accept that shape of risk — and increasingly the law does not allow it. By the end of 2025, 44 of 55 African countries had data-protection laws in force, several of which require sensitive data to stay inside the country.
Digital Policy Alert — Data Protection in Africa Roundup 2025; national statutes including Nigeria NDPA 2023, Rwanda 058/2021, Kenya DPA 2019, Tanzania PDPA 2022, South Africa POPIA
Kenya has 21 medical insurers. Twenty of them lost money last year. The one that made a profit is the one that put AI into claims processing and fraud detection. Its group chief executive puts the 2025 saving at KSh 1.2 billion, and says pre-authorisation went from over an hour to under a minute. Nobody in that story trained a model. They built the system around one.
Group chief executive of the insurer — public remarks, Africa BFSI Week 2026
We are a senior engineering studio, not a reseller and not a research lab. We build production AI systems and the software around them — mobile, web, backend, integration — and we deploy them inside your perimeter, in your country, under your rules. We use the strongest available models as interchangeable parts. We never let any one of them become a foundation you cannot remove.
What we build
One idea sits under everything we deliver. Several AI vendors do the computing. None of them ever sees enough to reconstruct your business.
Your perimeter — on-premise or in-country
Your data
customers · claims · documents
Keys and identities
never leave this box
The router — yours, not a vendor’s
splits each task · strips identifiers · picks the model · rejoins the answer
Private models on your own hardware
open-weight models for anything touching personal data
Only fragments leave the perimeter
Vendor A
sees fragment 1 · no identifiers · no context
Vendor B
sees fragment 2 · cannot join it to fragment 1
Vendor C — the checker
re-checks the answers of A and B
Answers come back and are rejoined inside the perimeter.
Every task is broken into parts before it leaves your network. Each provider receives only the fragment it needs. Names, account numbers and the logic that joins the parts back together stay inside. A leak at any one vendor exposes a fragment, not a customer.
A model-neutral layer picks the provider for each task by cost, quality and legal residency. Replacing a vendor is a configuration change, not a rebuild. Prices fall, providers fail, rules change — your system keeps running.
Anything touching personal or regulated data runs on open-weight models on your own hardware or in-country cloud. External providers are used only where sending data out is safe and lawful.
A fast, low-cost model does the volume. A stronger, independent model checks the result. Quality stays high and cost stays low, and two different vendors have to be wrong in the same way before a bad answer reaches a person.
On top of that platform
Large global companies no longer buy AI tools. They build an internal layer that connects their documents, systems and processes, and then run narrow assistants on top of it. We build the same thing for African institutions and governments — with narrow agents that have a measurable job, not a general chatbot that impresses in a demo and dies in production.
Retrieval over your own knowledge, agent workflows and domain-specific testing. Each system ships with monitoring, a fall-back path and a human review screen.
Native iOS and Android, web, backend, and integration into core systems and payment rails. The bench that carries the AI into production.
We map the work end to end and rebuild it around the model, so AI removes real effort instead of being added on top of it.
Opportunity mapping, cost and accuracy budgets, data-governance rules, and a written AI policy your regulator can read.
Hands-on training for the people who will use the system daily, and for the managers who have to trust its output.
On-premise, in-country cloud or hybrid, designed around each market’s data-protection law. Your data stays where the regulator expects it.
People and proof
A centaur team is a person and a machine working as one unit, where each does what it is good at. This is not a slogan. It is one of the most carefully measured results in the field — and it only works when someone designs the line between the two.
Faster work, 12% more tasks completed and 40% higher quality, when professionals used AI inside the range it is good at.
Dell’Acqua et al. — Navigating the Jagged Technological Frontier, Organization Science, 2026 (Harvard Business School / BCG field experiment)
More likely to be wrong when the same people used it outside that range. Drawing the boundary is the whole job.
Same study
Productivity gain for junior staff in a field study of 5,179 support agents — against +14% on average. AI lifts the inexperienced most.
Brynjolfsson, Li & Raymond — Generative AI at Work
Nigeria, South Africa and Kenya together hold well over a million developers, but the pool is young and senior specialists are hired away offshore. A centaur design turns a junior team into a mid-level team — the quickest way to close a gap hiring cannot.
GitHub Octoverse, 2024
People forgive human mistakes and refuse to forgive machine mistakes. We have watched organisations stop a working pilot because the model was wrong 0.5% of the time, while the people it supported were wrong around 6.5%. Twelve times better, and still rejected. The barrier is psychological, not technical. So we design for trust first: explainable output, a human who signs the decision, a clear audit trail, and an easy way to overrule the machine. Systems built this way survive their first bad week.
AppCraft delivery experience
Why AppCraft
Working together since 2011. Over 400 products delivered across four continents. More than 100 in-house engineers across design, backend, machine learning and security. Five ventures past $2m in revenue, one acquired by a tier-one telecom operator in 2025.
We ship our own AI products, not only client work. Our forged-image detector runs live in the browser. Voice-clone and forged-document detectors are built and are shown live on a call.
Banks, insurers, telecom operators and government bodies across the CIS, the Gulf and South-East Asia — markets with strict supervision, data-residency rules and on-premise requirements. The same discipline African regulators now ask for. Headquartered in Tbilisi, Georgia.
We read the regulation before we write the proposal. Deployment is designed around the data-protection law of your market, not retro-fitted to it after the audit.
How we start
Two to three hours with the people who own the process. We find where AI actually pays, and where it does not.
One process, in your environment, on your hardware. Designed to win an internal reference, not to bill a large first invoice.
What we agreed, what we did not, and what changes before the next conversation. No fine print.
Send a short brief, or write directly. We answer with what we would build, what we would not, and what it depends on.
Defending the same institution against AI-driven attack is the other half of this work — see the security page.
Sources. Africa Data Centres Association — Data Centres in Africa 2026 · Epoch AI — frontier model training-cost estimates · GSMA — AI for Africa · Digital Policy Alert — Data Protection in Africa Roundup 2025 · public remarks by a Kenyan insurer’s group chief executive, Africa BFSI Week 2026 · Dell’Acqua et al. — Navigating the Jagged Technological Frontier, Organization Science 2026 · Brynjolfsson, Li & Raymond — Generative AI at Work · GitHub Octoverse 2024. Figures are drawn from the public sources listed and are current to August 2026.