Project summary

This project grew from a real deployment we built and tested a farming chatbot with smallholder farmers and found serious failures misunderstanding local language, giving unsafe pesticide advice, and failing to communicate uncertainty. A separate incident involving an AI-generated agricultural recommendation in China and destroy 25 acres of the farm, surfaced through Polymarket, this reinforced the broader problem confident AI advice can be dangerously wrong in real farming contexts.

Our research asks: how can AI systems become safer and more reliable for high-stakes agricultural information?

We are building an agricultural Q&A dataset and evaluation framework to test model accuracy, safety, uncertainty, and failure modes. We will establish baselines across frontier and open-source models, then conduct targeted fine-tuning and compare performance before and after training.

What are this project's goals? How will you achieve them?

Project goals;

1. Firstly we will establish a reliable baseline for how current language models perform on agricultural questions, including accuracy, safety, reasoning, and uncertainty.

2. And we will Identify failure modes where AI produces incorrect, misleading, overconfident, or potentially harmful agricultural recommendations.

3. To test whether targeted training improves reliability by fine-tuning models on expert-reviewed agricultural data and comparing them against their original versions.

4. And develop a reproducible evaluation benchmark that can measure agricultural AI safety across different models and scenarios.

5. Lastly we will Study how AI safety and reliability transfer across languages and contexts, particularly where training data and expert resources are limited
   

The outcome of this project will be a reproducible benchmark, experimental results, and research paper that can contribute to safer AI development for high-stakes domais.

We will first expand our agricultural dataset with human and domain expert input, we will then evaluate several open weight language models using a held out test set, establish baseline results, and fine-tune selected models using parameter-efficient methods. Finally, we will compare pre and post training performance using standardized safety, accuracy, hallucination, and uncertainty metrics.

How will this funding be used?

$30,000. Dataset creation & expert validation expand the agricultural dataset, human annotation

$30,000. Compute & model training Fine tuning.

$15,000. Safety benchmark & evaluation develop rigorous tests for accuracy, hallucination, harmful advice

$10,000. Research personnel support for 3 months.

$5,000. Field validation Collect real world agricultural scenarios

$5,000. Infrastructure & tools: Research software, data storage and AI assistant subscription

$5,000. Publication & dissemination Research documentation, open source release.

Who is on your team? What's your track record on similar projects?

This is me Muhammad Zakari – Founder & Lead Researcher software engineering student & Self taught AI safety researcher

Jin Wang – Research Advisor PhD candidate in Economics at the University of Arizona.

Muhammad Khalilullah Uthman – Policy Advisor

Technical Assistant to the Katsina State Governor on Engineering and Automation.

· Abdulrahman Uthman

Professional software developer with 6 years experience and open-source contributor with over 4,000 GitHub commits.

Track Record on Similar Projects have already built and openly released the core tools for this research a 1,000-entry multilingual agricultural safety dataset, evaluation scripts that score models on safety and dialect accuracy, and a QLoRA fine-tuning pipeline. We also trained a proof of concept adapter and published it on Hugging Face, demonstrating that our pipeline works end to end.

In addition, build & tested a earlier agromind.chat with more than 100 farmers, documented the failures of frontier models, and used that evidence to shape the current benchmark. The work has been recognised through an $5k Tinker research grant, selection for the Africa Impact Challenge 2026 cohort, and selection for the United Nations Institute for Training and Research (UNITAR) Empowering Youth and Women programme in 2026 (ongoing) and an invitation to interview at The Engine (MIT).
Website: https://agroguardai.com

Chatbot: https://agromind.chat

GitHub: https://github.com/agroguardaiOSS

Hugging Face: https://huggingface.co/AgroguardAI

What are the most likely causes and outcomes if this project fails?

The main risks are limited cloud compute, budget constraints, and dataset quality. Insufficient GPU access could delay fine-tuning and evaluation, while dialect errors, duplicates, and code-switching could introduce misleading patterns into the training data.

If the project falls short, we would still release the dataset, evaluation tools, and fine-tuning pipeline openly on GitHub and Hugging Face, along with our findings on unsafe agricultural AI behavior in local languages. The likely failure would therefore be slower or narrower execution, not a complete loss of the research value.

How much money have you raised in the last 12 months, and from where?

I have raised $6,000 in the last 12 months this includes a $5,000 research grant from Thinking Machines and a $1,000 grant from an individual researcher secured through a cold email there's no institutional funding or venture capital has been raised so far