AgriTriage AI

Inspiration

Across the African continent, it is no news or mystery that rural farmers face massive agricultural challenges, ranging from rapidly spreading pests like Fall Armyworm to viral infections like the Cassava Mosaic Disease. Unfortunately, agricultural extension officers are chronically understaffed, and most farmers operate in regions where the internet network connection is poor or none at all. This creates an opportunity to bridge the gap by bringing AI directly to them.

What it does

AgriTriage AI is an entirely offline Retrieval-Augmented Generation (RAG) advisory system. By embedding standard agricultural manuals natively and running on budget consumer laptops, extension officers and farmers can receive instantaneous, highly accurate diagnostics and crop management strategies without needing any cloud access.

How we built it

We utilised a split-architecture approach to maximise performance on constrained hardware (i.e., an 8GB RAM laptop profile):

• AI & Backend: The llama.cpp was used to run a 4-bit quantized Llama-3.2-1B-Instruct-Q4_K_M model. The RAG pipeline was built in Python using FAISS for lightning-fast vector similarity search across a local agricultural database.
• Frontend UI: We stripped away bloated web frameworks and built a highly responsive desktop application using Tauri, React, and Vite.

Challenges we ran into

The biggest challenge was the strict 8GB RAM hardware limit imposed by the track. We initially explored larger 8B models but as expected, they consumed too much memory, leaving little room for the OS and our UI, risking Out-Of-Memory (OOM) crashes.

Accomplishments that we're proud of

We found the absolute perfect sweet spot with the Llama 3.2 1B model. We successfully crushed the baseline 15 Tokens Per Second requirement, achieving 17.5 TPS natively. Andd even better, our peak RAM usage (RSS) during inference is only 1.38 GB, meaning the app runs flawlessly alongside other programs on a budget laptop without any thermal throttling!

Built With

• faiss
• huggingface
• llama-3
• llama.cpp
• python
• react
• tauri
• ubuntu
• vite