Projects
Each project is designed with real-world deployment in mind — not just benchmark performance.
Selected Projects
Domain Adaptation in Medical AI
A medical AI research project focused on building models that generalize across hospitals, imaging devices, and patient populations when clinical data distributions differ between source and target sites.
Medical Report Generator
An LLM-powered system that produces structured clinical reports and physician-style patient summaries from diagnostic data, enhanced with RAG for evidence grounding.
Project LITE — Salmonella Early Detection
What it is: a team project built during the FLOuRISH Academic Entrepreneurship Global Program 2026 in Japan (Tokyo University of Agriculture and Technology), which went on to win the Interdisciplinary Innovation Award. LITE (Light-based Identification of Targeted Salmonella for Early Detection) is a rinsable sensor spray that farmers apply directly to tomato plants: DNA aptamers in the spray bind a Salmonella-specific surface protein, and that binding switches on a fluorescent signal visible under blue light where a contamination signal appears.
Why it matters: Salmonella on fresh produce is a growing, hard-to-see risk — a 2023-2025 outbreak linked to tomatoes affected hundreds of people across 17 European countries. Existing options force a trade-off: swab kits are fast but any positive still needs slow lab confirmation, and lab culture testing is reliable but too slow and costly for frequent, farm-wide checks. Neither supports the kind of continuous, localized monitoring that could catch contamination early and protect both consumers and farmers' livelihoods.
How it works: spray, recognition, fluorescence activation, automated detection, response and removal. The spray targets contamination directly on the plant surface, turning a slow, random-sample lab workflow into on-site "screen broadly, then confirm selectively" testing that can be repeated throughout the growing season with minimal equipment.
My part (AI): I designed the image-analysis workflow that makes the fluorescence signal faster and more consistent to read instead of relying only on human visual judgment. It takes a baseline photo of a plant before spraying and a second photo after, detects the plant/fruit regions, isolates fluorescence hotspots between the two images, computes a 0-100% signal score, and outputs a decision: PASS, HOLD (send to lab), or ISOLATE (remove).
Matokeo Yangu
A bilingual platform that helps Tanzanian students check exam results and receive AI-guided advice on university and career pathways.
WBC Detection & Classification
An automated pipeline for white blood cell detection, counting, and classification using YOLO and Vision Transformers. Optimized for deployment in resource-constrained clinical laboratories.
