Notes from Nyanga: the AI4I boot camp

I'm writing this from Tamarind Resort in Nyanga, a few days into the residential boot camp for the AI for Impact (AI4I) Challenge 2026. Inzwi is here on the Data track.

Some context first. POTRAZ launched AI4I on 29 June, together with the Ministry of ICT, Postal and Courier Services, the ITU and the Zimbabwe Centre for High Performance Computing. It's an AI-focused offshoot of POTRAZ's yearly Innovation Drive, and it lands in the same year as the National AI Strategy. Thirty teams made the shortlist: ten on Data, five on Design, ten on Development and five on Deployment.

Fine Touch is running the boot camp. They're the team behind a lot of POTRAZ's Start Your Business programmes, and you can tell. The week has a clear shape: sessions first, then two days of pitches (there are a lot of us), then the awards on the last day. Being out in the Eastern Highlands helps too. It's hard to be anywhere else in your head when you're this far from your usual desk.

Four tracks, not one

The thing I like most about AI4I is that it refuses to treat "AI" as a single activity. Each track is a different job.

Data is about collecting, curating and governing sovereign data. Design turns data into tools people can actually use and decide things with. Development builds the products, the prototypes and the MVPs. Deployment is the part everyone forgets until it bites: governance, bias testing, security and compliance once the thing is out in the world and people depend on it.

Most AI conversations skip straight to Development, because that's the part with demos. For a language like Shona or Ndebele you can't skip anything, because there's nothing to develop on. The data isn't there, and nobody outside Zimbabwe is going to come and get it for us. It isn't anyone else's priority, and there's no reason it should be.

Inzwi touches all four tracks eventually. But it belongs on the first one, and I want to explain why, because it's the most important thing I've understood about the project.

Where Inzwi started

Inzwi (the word means "voice") began as my final-year research project at MSU: a crowdsourcing platform for English and Shona data. The research question was modest. Could you get ordinary speakers to build a dataset for their own language, a sentence at a time, if you made it easy and treated them fairly?

The answer was yes, and it's still the heart of the project. What's changed is the scale of the ambition around it. Inzwi is now a speech platform, with recognition, synthesis and understanding for Shona and Ndebele. But every one of those depends on the same thing the research project was about: data that didn't exist before someone went and collected it properly.

Why it's a data project first

It would be easier to pitch Inzwi as "a Shona voice assistant". People can picture that. But it isn't the honest version. Inzwi is a data project first, and the models are what the data makes possible later.

Here's what "there's no data" actually means. Speech recognition learns from recordings paired with exact transcripts, many hours of them, from many different voices. Speech synthesis needs clean, consistent recordings of text read aloud. For English there are thousands of hours of both, free to download. For Shona and Ndebele there's almost nothing, and the little that exists leaves out most of how people really talk.

Shona isn't one way of speaking. Zezuru, Karanga, Manyika, Korekore and Ndau are different enough that a model trained mostly on one quietly fails people who speak another. It doesn't crash, it just gets things wrong more often for some people than for others, which is worse, because nobody notices. So Inzwi tracks each variety separately. Ndebele gets the same treatment, as its own Nguni language rather than something folded in beside Shona.

Then there's code-switching. Real Zimbabwean speech moves between languages mid-sentence: "Ndiri kuenda ku meeting manje manje, ndichaita call back ndadzoka." A dataset built from tidy, single-language sentences produces a model that falls over on exactly the way people speak every day. So the corpus keeps code-switching in, instead of cleaning it out.

How the corpus gets built

Collection happens at collect.inzwi.app, and it's deliberately split into small jobs that each take a minute. Speakers translate English sentences into natural Shona. Other speakers validate those translations. Validated sentences get recorded aloud. And in free speech, people answer questions naturally while others transcribe what they said, which is where most of the real, unscripted language comes from.

Around that sit the decisions that make it a sovereign dataset rather than just a dataset:

• Contributors are Zimbabwean speakers, and they're paid for their time.
• Consent is explicit, and it can be taken back.
• Everything is checked twice: by other speakers, and by automated audio checks, before it goes anywhere near training.
• The data is governed here, under the Cyber and Data Protection Act, and it stays here.

"Sovereign data" can sound like a slogan. For Inzwi it's that list: decisions about consent, pay, storage and labelling, made before a single model is trained, and very hard to retrofit afterwards.

What's changed in how I explain it

Getting ready to pitch has changed how I talk about Inzwi more than what Inzwi is.

A data project has to earn its place before its best demo exists. So the pitch can't lean on a model. It has to be about what the data makes possible: speech recognition that works for a Karanga speaker as well as a Zezuru one, a voice that reads out a clinic's instructions in Ndebele, systems that understand a sentence that's half Shona and half English.

Inzwi also has three audiences, and they need three different explanations. The people who record need to know they're treated fairly and why their dialect matters. The people who'll build on it need to know what it can do. And the people who'll eventually talk to whatever gets built need none of this. They just need it to work in their language.

And the parts I used to leave for the end, like consent, payment and governance, are the pitch on this track. They go first.

Pitches start soon. I'll write again when it's over.