category · ai & productivity
PrototypingOne job, done well
Assistants that do a single thing properly instead of a hundred things badly — and where they are wrong, a person finds out before a customer does.
Photo: Matt Biddulph on Wikimedia Commons, CC BY-SA 2.0
Projects & opportunities
3 projects across 3 subcategories. The rest are on our roadmap, with what we plan to explore.
Science
Elements, numbers and how things work.
Language tech
Software that speaks African languages.
Most African languages are barely served by software, from keyboards to speech recognition.
- Local-language tools
- Speech and voice
- Translation
Computer vision
Software that reads the world through a camera.
What we'd build
Four ideas, in the order we would test them. None are public.
-
01
Field notes to text
Offline transcription for interviews and site visits, in the language actually being spoken, so a conversation on a farm is not lost to a weak signal.
-
02
Documents to tables
Invoices, results and receipts turned into a usable sheet — and every extracted column traceable back to the page it came from.
-
03
A research assistant
Read a pile of sources and say what they support — and, just as importantly, what they do not cover.
-
04
A human in the loop
No output leaves the lab unreviewed. Every tool states its sources, and the person who checks it owns the result.
Where it stands
Prototyping
The first helpers are in build and used on our own work, which is the only honest test available to us: if a tool does not save us time, it does not deserve anyone else's.
This is the category where we expect to be most tempted to overstate. The rule is that we ship the boring version that works over the impressive version that demos well.