Identified Use Cases
Seven use cases identified for the node, each with sector, model profile, GPU intensity, and the Hub's role.
Seven use cases have been identified and mapped to the node. Each is presented with its sector, model profile, GPU intensity, and the Hub's role.
1. AI-Enabled Agricultural Advisory
Agriculture Moderate GPUMultilingual chatbot (WhatsApp, Telegram, IVR) for ~38 million smallholder farmers, in Hausa / Yoruba / Igbo. Target model N-ATLaS-LLM 8B (open-source Nigerian-language model). FarmerChat pilot completed in Delta State under the LIFE-ND Project.
2. AI-Based Oral Reading Fluency (ORF) Assessment
Education Moderate GPUReading-fluency assessment in local languages for ~86 million early-grade learners. Primary workload is ASR (Whisper-large-v3, ~1.5 GB VRAM) plus a lightweight scoring / feedback LLM; ~8–15 GB per active pipeline. Typically batch / asynchronous.
3. AI-Based Liquidity Management for Agents
Financial inclusion Low GPUForecasting cash / e-float needs for ~2 million mobile-money agents. Core models are lightweight (LSTM / XGBoost / time-series); an optional LLM chatbot adds advisory-style inference.
4. Self-CAIRE – Digital Health Coaching
Health Highest safety criticality Moderate–high GPULLM-powered health coach in the mDoc "Kem" platform for maternal health and chronic disease, via WhatsApp and voice – vision to serve 150,000+ members (current pilot ~300 women). Assumed model: a 7–13B (cited up to ~33B) instruction-tuned open-source model replacing the current GPT-4 dependency.
5. AI Point-of-Care Ultrasound (AI POCUS)
Health High GPU (training)Portable handheld ultrasound with AI interpretation for frontline workers. Edge inference runs on the device (Lumify / Android); the Hub's role is model training / fine-tuning on Nigerian datasets plus a cloud API for higher-complexity cases.
6. Malaria Intervention Allocation Copilot (ChatMRPT)
Health Low–moderate GPUDecision-support for malaria programme managers across a 218M at-risk population. Hybrid: geospatial ML (gradient boosting, spatial regression) + a conversational LLM layer (7–8B). Low user concurrency.
7. AI-Enabled Health Record Digitisation
Health Moderate GPU (throughput)OCR + NLP conversion of paper records across ~31,000–33,000 facilities into DHIS2 / EHR-compatible formats. Models: TrOCR / PaddleOCR-class OCR + lightweight NLP extraction. Throughput-bound batch workload (an H200 processes thousands of pages/hour).
Local Commerce
Reference architecture for a product that helps local traders list and sell their produce – photo-to-listing, price advice, and orders with payments and alerts, powered by the NAISH Hub.
Archive Overview
Gates-funded and related development work relevant to NAISH – the prior deployments, platforms, and use cases the Hub builds on.