farmpool

Inspiration

Small farms often produce high-quality crops but cannot individually supply the quantities required by supermarkets, restaurants and other large buyers. Buyers must also contact and coordinate with many farms separately, making procurement complicated, expensive and unreliable.

We created FarmPool, an AI-powered agriculture network that helps compatible farms combine their harvests into one reliable pooled order.

What it does

FarmPool supports both farmers and buyers.

Farmers can describe their harvest using a voice note instead of completing a complicated form manually. ElevenLabs transcribes the recording, while Claude extracts information such as the crop, variety, quantity, location, harvest date and price. When information is unclear, the AI asks a follow-up question instead of guessing.

Farmers can also upload a produce photo. Claude checks whether the image contains relevant produce and assesses its visible condition. Information is clearly labelled as self-reported, AI-screened or physically verified so buyers understand how reliable each claim is.

Buyers enter their required crop, variety, quantity, delivery date, minimum condition and maximum price. FarmPool removes incompatible harvests, generates possible farm combinations and uses a machine-learning model to rank the valid options by their likelihood of successfully fulfilling the order.

The buyer can review the selected farms, allocated quantities, available evidence, transport route, delivered cost and reasons why other farms were excluded. An AI coordinator explains the recommendation, while the buyer makes the final decision.

How we built it

We built the mobile application using React Native, Expo, Expo Router and TypeScript.

The backend was developed using Python and FastAPI. We integrated:

ElevenLabs: for speech-to-text Anthropic Claude: for structured voice extraction, clarification questions, photo screening and recommendation explanations Scikit-learn: for training a logistic-regression model that ranks feasible farm combinations Deterministic validation for quantities, dates, crop compatibility, pricing, transport costs and matching requirements

This creates a layered decision pipeline. Generative AI understands unstructured information, deterministic code enforces important constraints, machine learning ranks valid options, and people approve the final transaction.

Challenges we ran into

One major challenge was preventing unreliable information. Farmers could enter invalid dates, unrealistic quantities or unrelated crop varieties. We addressed this using structured crop catalogues, range checks, typed fields and confidence-based AI clarification.

Another challenge was assessing produce quality fairly. A photograph cannot reliably measure sweetness, internal condition or food safety. Instead of overstating the AI’s capabilities, FarmPool limits photo assessment to visible characteristics and clearly labels information that is still self-reported or unmeasured.

We also needed to ensure that AI recommendations could not bypass important business and safety rules. AI can interpret, explain and suggest, but deterministic code continues to control eligibility, quantities, routes and costs.

Finally, combining voice processing, image analysis, validation, matching and ranking into one mobile workflow required careful coordination between the Expo application and FastAPI backend.

Accomplishments that we're proud of

We are proud that FarmPool is more than a simple form or chatbot. It provides a complete farmer-to-buyer workflow in which each AI component contributes to the next stage.

Our prototype can convert a natural voice description into an editable harvest listing, screen produce photographs, validate farmer information, combine multiple farms, rank feasible pools and explain the final recommendation to a buyer.

We are also proud of making the system transparent. Buyers can see why farms were selected or excluded, how transport costs were calculated and what level of evidence supports each quality claim.

What we learned

We learned that AI is most effective when it handles uncertainty, natural language and complex evidence rather than replacing every rule or human decision.

Deterministic software remains essential for enforcing quantities, dates, costs and safety constraints. Human approval is also necessary for decisions involving payments, contracts, food safety and final orders.

We also learned that transparency is especially important in agriculture. Buyers need to know whether information is self-reported, AI-screened, document-supported or physically verified.

What's next for FarmPool

Our next steps are to add secure user and business verification, persistent cloud storage for photos and voice recordings, reviewer workflows and real farm and buyer data.

We also plan to develop a bounded recovery agent. If a farmer withdraws or fails a physical quality check, the agent could identify a replacement, recalculate the route and cost, and propose an updated pool. Any change would remain provisional until approved by the buyer.

In the longer term, FarmPool could integrate laboratory measurements, logistics providers, payment systems and live agricultural market data to become a trusted coordination platform for regional food supply.