Pitchfork

Homescreen
Homescreen - CSV file format + set constraints
Post Analysis - Table View
Post Analysis - Cards View

Inspiration

Inspired by firsthand observations in India, we built this project to use technology to promote fairer subsidy distribution and support small, underserved farms facing economic and technological challenges.

Problem Statement

About 85% of all U.S. farms are small-scale operations. Over 75% of these small farms receive little to no funding. Subsidies handed out by governments currently favor sheer acreage and volume, favoring large producers, leaving small-operation farmers financially vulnerable, unable to afford climate tech or long-term resilience.

What our product does

PitchFork is targeted towards the 90,000+ local governments around the world that decide subsidies to divide amongst their farms. Through various stat inputs about the farms, along with a total spending budget, and intuitive fairness constraint inputs, Pitchfork gathers and feeds 11 attributes of each farm, gathered from the location of the farm, into a trained ML model which outputs a predicted yield (t/ha) and climate risk of each farm. Finally, a mathematical, constraint-based subsidy of the total budget is allocated to each farm based on the yield/risk and constraints set by the user, and is displayed to the user through a sleek front-end display.

How we built it

We used Python to develop a Flask back-end that acts as the core pipeline for our project. We split this pipeline into individual scripts that handle different areas of our code. We created calls to Open-Meteo API to fetch location statistics, trained a custom PyTorch neural network to calculate yield & risk score using realistic generated data (AMD GPU usage reduced training time by 20 minutes), implemented a constraint-based optimization engine that translates policy fairness requirements into mathematical constraints and allocates subsidies accordingly, and finally used the Groq API with Llama 3.1 to generate our explanations of results for each farm. Our front-end was built using Next.js and hosted using Vercel.

Challenges we ran into

Training our model off of the realistic generated data was a challenge for us. Through training sessions and code improvements, and training with more diversified datasets, we were eventually able to get the model to predict realistic yield values and climate risks. Another challenge was managing our use of API calls to Open-Meteo API. We ran out of API calls to Open-Mateo during testing so our progress with troubleshooting was slowed during production.

Accomplishments that we're proud of

We incorporated our own trained ML model off of realistic data that can create predicted values for yield and climate risk. accurately determine the proper equity for farm income division, with strict fairness constraints. We created a working product that has full start to finish functionality with a polished front-end.

What we learned

We learned how to train machine-learning models and implement custom neural networks, while better grasping how to bridge the connection from front-end to back-end. We also gained more insight into the problem at hand and better understand the possible solutions.

What's next for Pitchfork

Pitchfork's value comes from its quantification of future potential climate risks and predicted yield to come up with objectively fair subsidies for each farm. Our next step will be to refine this process, implementing multiple ML models and comparing their results (Prize money will help fund training of these models).