Intelligence across the stack.

Most products bolt AI on in one place. A chat widget in the corner, a "generate with AI" button on one form, and a press release. It is easy to build and easy to demo, and it rarely changes how anyone works.

We took the other approach, which is slower and considerably less photogenic: work out where intelligence actually belongs in each workflow, and put it there. This issue is the inventory.

AfriKDP: twenty-two capabilities, and a studio

AfriKDP is our most demanding environment, because publishing is not one job. It is writing, editing, designing, pricing, distributing, marketing, selling, and — on the other side — discovering, reading, studying and learning. Each of those is a different problem.

So AfriKDP AI and AfriKDP Copilot now run across the whole arc:

• For writing — manuscript assistance, consistency checking, translation and quality review, localisation.
• For preparing a book — automatic book analysis on upload, chapter overviews, metadata generation you review before applying, cover and creative concepts, audiobook preparation.
• For selling — marketing studio, marketing calendar, audience builder, promotion assistant, buyer analysis, author insights, marketplace intelligence.
• For readers — book summaries, Ask the Book with retrieval that cites the chapter it answered from, reading assistance, comparison, discovery by description, recommendations, quizzes, study mode and a tutor.

Translation deserves a note. We deliberately block automated publication of translations into Yorùbá, Igbo, Hausa, Amharic and Zulu, and require human review first. Our own research says frontier models are weakest exactly there — so shipping unreviewed machine output in those languages would be both bad practice and self-refuting.

The Book Writer Studio

The hardest thing we built this cycle. A full-length book does not fit in a model's context window, so we did not try to make it fit.

The brief's page count becomes a plan: words per page for the chosen trim size, then chapters, then sections. Each section is one call, carrying a running summary of what has been covered and the tail of the previous section.

Continuity holds, and no single call is ever oversized. A 54-page 6×9 book is roughly 15,000 words, five chapters, sixteen calls. A 500-page book is the same process with more of them.

The second design decision matters as much as the first: the model writes structure; our code renders the file. Sections come back as markdown with image slots marked in place, and our renderers turn that into a properly typeset artefact — a DOCX with theme fonts, real page sizes, running headers, page numbers, a genuine Word table-of-contents field, chapter openers, drop caps, figures with captions and full front and back matter; a PDF with the same design and real page numbers; a valid EPUB 3 with cover and navigation; or plain Markdown.

Authors then work on a page canvas in their chosen theme, with AI available on any selection — rewrite, expand, shorten, simplify, make vivid, fix grammar, shift register, convert to dialogue or bullets, continue, translate. Images can be generated, regenerated, uploaded, captioned, moved or deleted. There are generators for dedications, forewords, introductions, conclusions, author bios, acknowledgements, glossaries and further reading. There are snapshots with restore, autosave, live word and page estimates, and a design panel with live preview.

And then there is a Publish button, which creates a draft AfriKDP book — manuscript and cover — that the author prices and sells like any other product. The whole chain, from a one-line brief to a book on sale, stays inside one platform.

AfriAssist: six agents, not one model

AfriAssist is live at afriassist.com, and it is built on a different architectural idea: cooperating agents rather than one model answering everything.

A user states a goal in plain language — "I am launching my new fashion collection". Strategy turns that into a content plan. Content writes the posts. Creative designs them. Publishing and Engagement puts them out and watches what happens. Memory carries what worked into the next cycle. A Video agent is in development.

The capability surface spans AI assistance, content creation, image generation, video creation, music and audio, and automation — with the connections to publish where these businesses actually are, and to monitor the channels they actually answer on.

One principle we will not be relaxing: AI drafts and plans autonomously, but publishing requires explicit approval unless the user deliberately opts into auto-publish. An assistant that posts to your audience without asking is not a feature.

Skilluxo: the intelligent layer

Skilluxo is live, with escrow protecting every order, verified freelancers across design, writing, development and marketing, direct messaging and live calls between buyers and sellers, and tiered packages. Free to join; a fee only when you get paid.

Skilluxo AI is the next layer, and our intent is that it runs across the whole professional services journey rather than sitting in one screen: discovering opportunities, building profiles, matching skills to jobs, drafting and improving proposals, clarifying requirements, supporting delivery, teaching new skills, and automating the repetitive parts of running a freelance business.

Why it is all routable

Every one of these features runs behind the Omnis Gateway. Interactive features answer from whichever backend is fastest; background jobs run on Omnis. One environment variable moves a single feature between backends; another moves everything at once. Neither requires a code change or a deploy.

That is what lets us treat model quality as a measurement rather than a belief. Our admin tooling reports requests, failures, truncation and fallbacks per provider, per feature — so the question "is our own model good enough for this yet?" has an answer in a table rather than an argument in a meeting.

Twenty-two capabilities shipped is the visible result. Twenty-two capabilities that can each switch to our own intelligence independently is the actual asset.

The honest scoreboard

• Shipped and live: AfriKDP AI and Copilot, the Book Writer Studio, AfriAssist with its agent system, Skilluxo with escrow and live communication.
• Running on adapted open-weight models today: most interactive AI features. We are not pretending otherwise.
• Running on our own models today: background and batch workloads, where our adapters have cleared the bar.
• Not yet built: Skilluxo AI as described above. It is specified and scheduled, not shipped.

Next issue: the Omnis ladder, the first proprietary model we expect to be genuinely shippable, opening our infrastructure to outside developers, and what capital accelerates.