Applied AI research,
grounded in Africa.
GCT Exchange is a network of African researchers, clinicians, agronomists, and engineers building open AI tools for the problems closest to home.
- trainingKinyarwanda ASR · v0.4KIN-07 · 67%
- evaluationCassava leaf disease classifierAGR-12 · F1 0.91
- field testAntenatal risk model · pilotHLT-03 · n=820
- releasedSahel drought forecast (10d)CLM-19 · v1.2
Four areas where we focus our work.
African Languages
Open speech, translation, and NLP models for under-resourced languages across the continent.
Smallholder Agriculture
Computer vision and forecasting tools that help farmers respond to pests, weather, and yield.
Public Health
Diagnostic support and outbreak modeling co-designed with frontline clinicians.
Climate & Land
Remote sensing for drought, deforestation, and water resilience in the Sahel and beyond.
A full-stack research practice.
Foundation models
Pretraining and fine-tuning of speech, vision, and language models on Africa-relevant corpora.
Open datasets
Consent-based, ethically sourced datasets with documented provenance and permissive licenses.
Edge inference
Quantization and on-device deployment for entry-level Android, microcontrollers, and field hardware.
Reproducible pipelines
Versioned data, code, and weights so every result is verifiable and re-runnable.
Field evaluation
We measure model utility where it lands — clinics, farms, classrooms — not just on benchmarks.
Responsible release
Model cards, dual-use review, and community feedback loops before public release.
A short, honest history.
Seven researchers across four countries form the initial collective.
Kinyarwanda ASR v0.1 and cassava disease v1 ship under CC-BY-SA.
First cohort of 12 fellows — 70% women, 100% Africa-based.
Models deployed across 140 clinics and 900+ farms in pilot programs.
The model behind the latest release.
Kinyarwanda ASR v0.4
Amina Mukamana · Director, Languages
Speech recognition for Kinyarwanda — built on 1,200 hours of community-contributed audio, with a semi-supervised pretraining pass that cut word error by 41% over our previous baseline.
From conversation to deployment, in four steps.
Listen
We start with the people closest to the problem — clinicians, farmers, teachers, linguists.
Co-design
Researchers and practitioners scope the work together. Constraints come from the field, not the lab.
Build & evaluate
We train, fine-tune, and benchmark — then test in the environment the model will actually run in.
Release & iterate
Open weights, datasets, and model cards ship together. Community feedback feeds the next version.
Across nine countries — and growing.
What collaborators say.
"GCT's open release of the Kinyarwanda model let us ship a clinic voice intake tool in three weeks — work that would have taken us a year alone."
"Their fellowship gave me compute, mentorship, and a real co-author. I left with a paper, a deployed model, and a network I still rely on."
"What stands out is the discipline. Honest scope, open releases, and field evaluation that actually means something."
Latest from the network.
What we got wrong about Kinyarwanda tone in v0.3
A retrospective on tone modeling and the dataset patch that fixed our F1 collapse.
Notes from a week with cassava farmers in the Ashanti region
Field deployments humble models fast. What 14 farmers taught us about UX, latency, and trust.
Why our antenatal model intentionally returns 'unsure'
Calibration matters more than accuracy in clinical decision support.