Technology aimed at problems that nobody else is being paid to solve — and, in time, the work of our learners.
Student work
No subject has opened yet, so there are no graduates and no student projects. We would rather show you an empty page than invent one. Here is exactly what will appear here, and how to be first on it.
Technology for Afghanistan
No global company has a commercial reason to solve these. That is precisely why they are ours. Each one is a real engineering problem, not a slogan.
Speech recognition for the languages spoken here is poor because there is almost no openly licensed audio to train on. We are collecting read speech from volunteers — thirty seconds each, from a phone — with documented consent and a permissive licence, so that anyone can build on it.
What it unlocks: dictation, voice interfaces for people who cannot read fluently, subtitling, and call-centre work that currently goes elsewhere.
Afghan school textbooks exist mostly as scans and photographs, and existing recognisers handle the printed script here badly. We are building a corrected page corpus so that a usable recogniser can be trained.
What it unlocks: searchable textbooks, screen readers for blind students, and every retrieval system that needs the curriculum as text.
Wheat rust, and diseases of grape, pomegranate and almond. A model that works from a phone photograph in poor light, runs without a connection, and says "I am not sure, ask a person" instead of guessing.
What it unlocks: earlier treatment for farms far from any agricultural extension officer.
Comparing satellite imagery before and after a quake to rank which settlements are worst hit, in the hours when responders must choose where to go first. Built openly so any agency can use it.
What it unlocks: faster triage of response, using imagery that is already public and free.
Extra practice problems for every subject, with worked solutions — the thing learners always run out of first. Written by whoever has just finished that subject, which is when the confusing parts are still fresh.
What it unlocks: teachers anywhere can set homework from it, and learners can practise until it sticks.
The whole curriculum in an Android app that syncs when there is a connection and works entirely without one. Under 50 MB, and it must run on a four-year-old phone.
What it unlocks: study for learners whose data allowance runs out on day nine of the month.
Further out
These depend on the datasets above. We list them so the sequence is clear — and so nobody mistakes the ambition for a promise.
A patient tutor that explains a lesson again, differently, at two in the morning. Needs the speech corpus and the textbook corpus first.
Plain-language answers to common health questions, grounded in sources a clinician approved, and refusing anything that needs a doctor.
Explaining official forms and procedures in the language the applicant speaks, so a form is not a barrier to a service someone is entitled to.
Localised drought and flood signals for farming districts, from open climate data, delivered by SMS where there is no smartphone.
Screen reading and dictation for the languages spoken here, which barely exist today. This is what the speech corpus is ultimately for.
Afghan-authored papers on low-resource language technology, in the venues where the field reads. Slow, and worth it.
We would rather publish a short list of finished things than a long list of intentions. If a project above stalls, we will say so on this page rather than quietly deleting it.