Applied work on how evidence and data can help development programs reach more people: an offline AI tool for coffee cooperatives in Kenya, an evaluation for CARE, and a study of how a mental health program scaled in Zimbabwe.
Majani Relay: Offline AI for Coffee Cooperatives in Kenya
with Yicong Li
Runner-up, ****World Bank "Small AI for Development" Hackathon, Agriculture track (October 2026)
- Coffee leaf rust is a fungus that spreads from tree to tree; one study in Uganda found it cut Arabica coffee income by about half. Catching it early mostly depends on an extension officer seeing the trees. In Kenya, one extension officer typically serves 1,500 to 3,000 farmers, and reliance on paper reporting delays information.
- Majani Relay is an offline phone app for coffee cooperatives. Farmers trained by their cooperative photograph 3 leaves on each of 5 trees on other members' farms. A small AI model on the phone names the leaf problem in Swahili text and audio, or says "not sure — the officer will look". The officer makes the final call, and those labels update the model on the phone. A co-op screen turns the photos into village rust counts and ranks villages, so the officer knows where to go first.
- On healthy and rust field photos from Ecuador, never used in training, the app answers 91% of photos, and 89% of those answers are right. On farm photos from Uganda it first sends 93% to the officer; after 100 officer labels it answers 46%, and 94% of those answers are right (dataset labels stand in for the officer). Next is a proposed 90-day pilot with a coffee cooperative in Kirinyaga, scored on a sealed set of Kenyan photos.
<aside>
👉
Live app (demo PIN 2026) · Code · One-page summary · Demo video · Technical video
</aside>
Evaluating Evidence for a Large-Scale International Development Program
Research Intern, System Level Study
CARE USA, Spring 2026
- Evaluated the implementation and impact of IPEL, a USAID-funded large-scale education program in India by combining econometric analysis, process tracing and Bayesian methods.
Grandmothers, Benches, and a Mental Health Revolution: Scaling Care in Zimbabwe
with Chanzé Ahsan
Course report, DEV 320M (Winter 2026)
- An analysis of how the Friendship Bench took problem-solving therapy delivered by trained lay grandmothers ("gogos") from early trials to national coverage across Zimbabwe's ten provinces. It argues that effectiveness depends less on providers' credentials than on structured design, supervision and fit with the health system, and recommends formal recognition of task-shared care and recurring budget lines for mental health in primary care. Report available on request.