FarmHealth AI

Inspiration FarmHealth AI

Inspiration

Millions of African farmers operate in environments with limited internet connectivity, unreliable electricity, and restricted access to agricultural extension services. While modern AI systems can provide valuable farming advice, most require cloud access, expensive hardware, or continuous internet connectivity.

FarmHealth AI was inspired by a simple question:

Can we build an agricultural AI assistant that runs entirely on the laptops and computers farmers, extension officers, and agricultural organizations already own?**

Our goal is to make practical agricultural knowledge available offline, helping users make better decisions about crops, livestock, weather risks, farm operations, and animal health.

What It Does

FarmHealth AI is an offline agricultural and livestock-health assistant designed for African farming environments.

The system helps users:

Diagnose common crop and livestock problems
Identify diseases and preventive measures
Understand weather-related farming decisions
Improve farm operations and storage practices
Access localized agricultural knowledge
Receive guidance in English and selected African languages

The project focuses on agriculture in Northern Nigeria, including regions such as Nasarawa, Kaduna, Adamawa, and the FCT agricultural belt.

How We Built It

FarmHealth AI is being built using a lightweight on-device AI architecture optimized for the Africa Deep Tech Challenge 2026 standard hardware profile.

Key design principles include:

Offline-first operation
Low-memory deployment
CPU-only inference
GGUF model format
llama.cpp runtime compatibility
Quantized language models optimized for 8 GB RAM systems

The solution combines:

A fine-tuned agriculture-focused language model
Structured agriculture and livestock-health datasets
Multiple-choice and free-text evaluation datasets
A future retrieval layer for localized agricultural knowledge

We are currently evaluating multiple foundation models including Qwen and Tiny Aya to determine the best balance between accuracy, speed, memory efficiency, and multilingual capability.

Challenges We Ran Into

Building AI for African deployment environments presents unique challenges:

Limited compute resources
Strict 8 GB RAM constraints
Offline operation requirements
Agricultural data collection and validation
Hausa language support and localization
Balancing model accuracy with performance

Another challenge has been designing datasets that reflect real agricultural situations rather than generic internet questions. We are focusing on practical scenarios that farmers and extension officers encounter in everyday operations.

Accomplishments That We're Proud Of

So far we have:

Defined a complete agricultural knowledge taxonomy
Created structured dataset specifications
Built an agriculture-focused MCQ evaluation framework
Designed a multilingual data strategy
Established a benchmarking framework aligned with ADTC evaluation criteria
Developed a model comparison framework for selecting the final deployment model

What We Learned

This project reinforced that successful AI systems are not only about model size. Dataset quality, evaluation methodology, optimization, and user relevance are equally important.

We also learned that solving African problems effectively requires designing for real-world constraints such as intermittent connectivity, older hardware, and language diversity.

What's Next for FarmHealth AI

Our next milestones include:

Expanding the agriculture dataset
Benchmarking candidate models
Fine-tuning the selected model
Building the offline retrieval system
Integrating localized agricultural knowledge sources
Completing performance profiling using the ADTC benchmark framework
Preparing for Gate 1 submission and evaluation

FarmHealth AI aims to demonstrate that useful, domain-specific AI can run effectively on the hardware already available across Africa, helping democratize access to agricultural knowledge without requiring constant internet access.

What it does How we built it Challenges we ran into Accomplishments that we're proud of What we learned What's next for FarmHealth AI