Tauron

Predicting livestock disease 48 hours before symptoms appear

Hackathon · Cornell Digital Ag Hackathon 2026 · Team of 6 · Built the ML pipeline and trained the model · February — March 2026

Overview

Tauron is an early warning system for dairy herd disease, built with a team of 6 at the Cornell Digital Ag Hackathon (February 27 to March 1, 2026). It predicts mastitis, bovine respiratory disease, and lameness risk 48 hours ahead and turns each prediction into a plain English alert.

My role

I wrote 64% of the commits. I built the ML pipeline, trained the model, integrated an external cattle disease dataset, and added symptom perturbation to the synthetic data, which moved AUROC from 0.50 to 0.995. I also built frontend pieces like CSV upload and prediction history, documented the explainability method, and made the brief and slides.

How it works

A GRU encodes a 7 day window of 9 sensor features per cow, two GraphSAGE layers pass information over the pen and bunk contact graph, and a sigmoid head outputs a risk for each disease. Explanations come from a single backward pass instead of GNNExplainer: about 5 ms per cow instead of about 200 ms, a 40x speedup, with matching top feature rankings on our data.

The top features go to Mistral 7B running locally through Ollama, with Claude and template fallbacks, to write the alert.

Status

Runs on a synthetic 60 cow herd with a FastAPI backend and a React and D3 herd map.