MAM-AI: An On-Device Medical Retrieval-Augmented Generation System for Nurses and Midwives in Zanzibar

A new medical question-answering system designed for nurses and midwives in Zanzibar runs entirely on a standard Android phone, operating without any internet connection to provide point-of-care guidance drawn from 87 clinical documents [1]. The system, called MAM-AI, is described in a paper submitted on 28 Jun 2026 to arXiv [1]. It targets a persistent challenge: maternal and newborn mortality in sub-Saharan Africa, where care is often provided by nurses who lack midwifery training to international standards and where consulting lengthy guidelines is complicated by intermittent connectivity [1][2]. The assistant embeds a user's question using a 300M-parameter model and matches it against a curated corpus of 63,650 passages, then generates an answer with citations via a 4B-parameter int4 generator [1]. No query data leaves the device [1]. The research team evaluated the deployed configuration with a layered methodology, scoring retriever, generator, end-to-end, and latency performance using LLM judges validated against physician rubrics [1]. The evaluation found that on-device retrieval is essentially solved: the 300M embedder ranked third of seven retrievers tested and rivaled cloud-based systems [1]. The bottleneck shifted to the small generator. Adding retrieved context did not improve its answers, and at 4B parameters the model could not be both helpful and safe simultaneously [1]. The more helpful of two same-size candidates produced genuine dangerous errors, so the team deployed the safer one, which was about twice as faithful to its sources [1]. A redesigned prompt cut the deflection rate from 33% to 3%, recovering helpfulness [1]. Corpus quality proved decisive. When the knowledge base contained the right passage, answers were specific and actionable; when it did not, responses became vague [1]. The authors describe MAM-AI as a thoroughly evaluated, open-source research prototype, not a fielded product, and have released the system, knowledge base, benchmarks, and evaluation harness [1][2]. The work lands in a landscape where on-device AI is gaining attention for its privacy and connectivity advantages. Other recent research has explored generative systems for specialized domains, though a review of quantum circuit generation tools found none had been evaluated end-to-end on actual quantum hardware, highlighting a broader gap between prototype and practical deployment [3]. Meanwhile, platforms like Hugging Face have been working to make machine learning research more accessible, integrating interactive demos directly onto arXiv abstract pages to allow users to test models without writing code [4][5].