The technology underneath every AfriStack product.
AfriStack is an AI and infrastructure company. Our products are where that technology gets deployed, measured and improved — not the other way round.
Four interconnected areas.
Artificial Intelligence
We research, train, adapt, evaluate and deploy AI models and intelligent systems designed for real-world applications — with our own tokenizer, architecture, training loop and evaluation harness.
Omnis AI →AI Infrastructure
The layer that runs, integrates, orchestrates and routes intelligence across every product — and that will eventually train our own models at scale.
How it works →Digital Platforms
Scalable platforms combining AI, automation, commerce, marketplaces and specialised workflows — AfriKDP, Skilluxo and AfriAssist are all live today.
See the products →Developer & Business Infrastructure
APIs, AI services and automation capabilities designed to be consumed by other businesses and developers — starting with the AfriKDP Publishing API, already live.
Developer infrastructure →One gateway. Swappable intelligence.
The single most important architectural decision we have made is also the least visible one.
No AfriStack product ever calls a model directly. Every product calls the Omnis Gateway, and the gateway decides which model answers.
That rule is what lets our own intelligence take over gradually. Swapping a backend becomes a configuration change rather than a rewrite, so Omnis can claim one capability at a time — as soon as it is measurably good enough for that capability, and not a moment before.
It also means our products never block on our research, and our research never blocks on our products. The two tracks run in parallel and meet at one interface.
Routing in production today
Inside AfriKDP, every AI feature is individually routable. Interactive features answer from whichever backend is fastest; background jobs run on Omnis. Two switches move a single feature or every feature at once, with no code change.
If a backend is unavailable, features fall back automatically rather than failing. The admin panel reports requests, failures, truncation and fallbacks per provider — which is how we know when Omnis is ready to take a feature back.
Inference
Self-hosted serving rather than renting someone else's API, so we control cost, latency and where the data goes.
Orchestration
Per-feature routing, automatic fallback, and measurement of every call so model quality is a number, not an opinion.
Data sovereignty
In-region inference keeps customer data inside the jurisdictions our enterprise and institutional customers answer to.
We are not just integrating AI. We are learning to build it.
Our AI journey began with a practical question: can a small technology company build and train its own AI systems with limited resources? We started with consumer hardware and built the pipeline ourselves.
- Data processing and dataset preparationCleaning, quality-scoring, deduplication and manifesting — with a source registry that nothing trains without an entry in.
- TokenizationOur own byte-level BPE tokenizer, with a multilingual compression report to check it treats African languages fairly.
- Architecture and pretrainingA decoder-only transformer with RMSNorm pre-normalisation, RoPE positional embeddings, SwiGLU feed-forward, tied embeddings and flash attention.
- Fine-tuning and LoRA adaptersInstruction tuning on our own pipeline, plus successive LoRA adapter generations over an adapted open-weight base.
- Evaluation and product benchmarksHeld-out perplexity, next-token accuracy and factual recall, measured against benchmarks derived from our own products rather than generic leaderboards.
- Inference, orchestration and deploymentServing the result behind the gateway, where every call is measured in production.
What we learned the expensive way
Every failure at the smallest model size costs hours. The same failure at frontier scale costs six figures. So we do not skip stages — each model in the ladder exists to make the next one cheaper to get right.
We are equally clear about what our early models cannot do. A 42M-parameter model is not a product; it is proof that the pipeline works end to end. Publishing the limitations alongside the results is the only way the results mean anything.
Where we intend to be unmatched
We do not intend to compete with frontier labs on general English reasoning. We intend to win where they are structurally weak and will stay weak.
From AI applications to AI infrastructure.
Our roadmap extends beyond using third-party AI models.
Integrate and adapt
Integrate and adapt the best available models behind our own gateway, while our from-scratch line proves the pipeline.
Specialised models
Develop specialised models and AI systems for our own products — translation, extraction, classification and narrow assistants.
Proprietary scale
Build larger proprietary models with specialised African and global language capabilities.
Intelligence services
Provide AI infrastructure, models, APIs and agents to businesses, developers, partners and enterprises.
How we think about cost
Training compute scales roughly with parameters multiplied by tokens, and the honest consequence is that matching a frontier general-purpose model head-on is a nine-figure problem before you count the research team. That door is closed — to us and to almost everyone.
The door that is open is a smaller model that is genuinely the best in the world at a narrow, valuable task. A model that is unmatched at Yorùbá is a far better business than one that is fortieth-best at English. That is the target we are actually aiming at, and our capital plan is built around reaching it — not around a headline parameter count.