Selected work and engineering evidence

Problem: Multilingual health information is hard to access reliably, especially on low-bandwidth connections and messaging platforms.

Built: A live multilingual AI health-information assistant on the web and Meta WhatsApp Cloud API, offering educational triage support, multimodal image understanding, vital-sign tracking and physician-ready consultation summaries.

Problem: Kenyan learners need curriculum-aligned lessons, assessments and affordable access that works on mobile and low-cost connectivity.

Built: A KICD-aligned CBC platform with strand lessons, term assessments, an AI tutor, WhatsApp reminders and quizzes, and bilingual English/Kiswahili narration.

Problem: Researchers spend hours manually critiquing methodology, auditing datasets and drafting grant proposals from academic papers.

Built: A five-agent research-paper intelligence platform — OCR, vision, critics, synthesis and grant — that critiques methodology, audits datasets, proposes experiments and drafts NSF/NIH-style grant outlines using Mistral models.

Problem: Residents need a clear way to submit structured feedback, and a constituency office needs aggregated insight without automating decisions that require human judgement.

Built: A bilingual resident-feedback platform that aggregates submissions and uses AI to suggest themes, with every AI suggestion reviewed by staff. PII is scrubbed before any AI call, and there are no automated eligibility decisions.

Problem: Blind and low-vision users need practical visual assistance that respects privacy and works when connectivity is limited.

Built: A Flutter Android prototype with offline OCR, speech and vision assistance, a Cloudflare Worker backend, and a TalkBack-first design that asks for consent before any image leaves the phone.

Problem: Viewers need personalised movie recommendations, and public-leaderboard scores do not translate directly to real-world KPIs.

Built: A hybrid content-based and collaborative-filtering recommender built on about 10 million ratings, using SVD matrix factorisation and LightGBM. Public-leaderboard RMSE improved from 1.21 to 0.82 across eight submissions.