Building a voice-enabled study bot revealed that "it worked in the demo" and "it actually works" are two very different things — and why rigorous evaluation is the most important part of building AI tools for high-stakes domains.
Fine-tuning GPT-4o-mini as a dedicated machine translation layer for Sheng — a Swahili-English slang spoken by young people across Kenya — to support a family planning chatbot. Presented at OpenAI DevDay 2024.
Training a random forest classifier to predict nutritional decline in 22,767 children across Madhya Pradesh and Odisha — and why identifying the right problem required as much collaboration as the modelling itself.
Applying ML-based record linkage to connect community health worker and clinic records in rural Malawi — and why precision and recall tradeoffs have real consequences for patient care.
Building computer vision tools to replace handwritten logbooks for health workers in Nigeria — and what it actually takes to make automation work in low-resource environments.
On what it means to be a responsible data scientist — and why the obligation doesn't end when the model performs well.
More case studies coming soon