The promise of AI in healthcare assumes infrastructure that much of the world doesn't have. Large language models running in data centers, retrieval systems querying cloud databases, applications that fail gracefully when connectivity drops - these architectures implicitly require reliable internet, modern devices, and users comfortable with digital interfaces.

What happens when none of those assumptions hold?

We're sharing insights from our collaboration with Kajou, a Bibliothèques Sans Frontières initiative, on kSanté, an AI assistant that confronts this question directly. The system delivers medical information to community health workers operating in conditions that break conventional deployment patterns.

The Problem Space

Community health workers in remote West African regions face a compounding set of constraints that make standard solutions inviable:

• Connectivity is absent, not poor: These aren't areas with slow 3G - they have no signal at all for days or weeks. Any system requiring network access, even intermittently, is nonfunctional.

• Medical knowledge gaps are structural: Workers lack formal medical training and have no familiarity with clinical terminology. They describe symptoms in everyday language, often in local languages that medical resources weren't written for.

• Hardware is what people actually have: Deployment targets are the aging Android devices workers already own - phones with 2-3GB RAM, limited storage, and processors from 2018. These aren't devices that can run typical AI workloads.

• Linguistic complexity compounds everything: Medical information exists in French, but workers and patients might have low knowledge of it. 

When a patient presents with symptoms, a health worker needs immediate access to verified treatment protocols. The current reality is stark: they have 30 PDF manuals that they can't search effectively. The information gap occurs precisely at the point of care.

Traditional digital health solutions simply don't function in this context. The infrastructure assumptions embedded in their design make them irrelevant. Yet the medical need is acute, and these workers represent the only point of care for populations that number in the millions.

This creates a design problem: how do you build an AI system when the standard architecture is impossible?

Technical Approach

The kSanté system addresses these constraints through architectural decisions that treat them as requirements, not obstacles to work around:

Natural Language Understanding in Low-Resource Languages

The system processes queries in conversational language rather than requiring medical terminology. A worker describing symptoms as "enfant a de la fièvre et refuse de manger" receives relevant information about pediatric malaria without needing to know diagnostic codes or clinical terms.

This presents significant technical challenges. Standard embedding models perform poorly on the informal, low-literacy French common in these regions - and even worse on Wolof. We've had to extensively work on these models to handle the linguistic reality of our users, not idealized textbook language.

Fully Local Inference

The entire AI pipeline runs on-device with no server dependency.

Meeting latency requirements for RAG workflows on constrained hardware required substantial optimization work. We've leveraged Pleias's research on efficient on-device inference to deliver acceptable response times on devices with as little as 3GB RAM.

Knowledge Base Coverage

The retrieval system indexes verified medical information spanning malaria protocols, infectious disease diagnostics, and treatment guidelines. Content is curated from national health ministry protocols (locked in rich-format PDFs, once again very difficult to parse…), processed and prepared for rich semantic retrieval.

Deployment Realities

Field testing has exposed the gap between controlled development environments and actual deployment conditions:

• Literacy variability: Users with very low literacy produce queries that are challenging for any NLU system to process effectively. This requires careful engineering and extensive error handling.

• Content maintenance: Medical protocols evolve, requiring a system for distributing knowledge base updates through channels other than app stores (which assume connectivity).

• 
  

Cost-Effectiveness for Public Health

For Ministries of Health operating on constrained budgets, kSanté represents a scalable approach to training and supporting community health workers. The system provides:

• Continuous access to medical information without recurring connectivity costs

• Self-directed learning opportunities for workers to review protocols and improve their knowledge

• Consistent information delivery across geographically dispersed teams

• A deployment model that works with existing hardware infrastructure

This makes AI-assisted medical information accessible at a fraction of the cost of traditional training programs or centralized call centers.

Open Questions

Building AI for these environments surfaces broader questions about deployment patterns:

How do we evaluate model performance when ground truth is difficult to establish and user feedback mechanisms are limited? What safety mechanisms are appropriate for medical information systems that operate without oversight? How do we balance model capability with inference constraints when both are critical?

The work on kSanté demonstrates that effective AI deployment in resource-constrained environments requires rethinking assumptions about infrastructure, language, and user interaction patterns.