Tauron

GNN-based early-warning system for dairy herd disease — 48-hour prediction horizon.

Cornell Digital Ag Hackathon · February 27 – March 1, 2026

Collaborators

• Aritro Ganguly
• Adi Prathapa
• Connor Mark
• Krish Jana
• Vikram Davey
• Suchit Basineni

---

Architecture

Sensor data (7-day window)
  ↓
GRU encoder (9 features → 128-d)
  ↓
GraphSAGE ×2 (pen & bunk contact graph)
  ↓
Sigmoid head → [mastitis, BRD, lameness] risk
  ↓
Gradient XAI → plain-English alert (Ollama / Mistral-7B)

Component	Stack
Model	PyTorch Geometric · GRU + SAGEConv
Backend	FastAPI · GET /herd · GET /explain/{cow_id}
Frontend	React (CDN) · D3.js force-directed herd map
LLM	Ollama (Mistral-7B) → Claude API fallback → template fallback

---

Explainability (XAI) Methodology

Our brief references GNNExplainer (Ying et al., 2019) as the explainability method. After implementation and benchmarking we chose gradient-based feature attribution instead. This section explains why.

What We Use

When a farmer clicks a cow in the Herd Map, Tauron runs a single backward pass through the dominant-disease output neuron. The resulting gradients on the 7-day sensor input tensor are:

1. Averaged over the time window (mean |∂risk / ∂x_t,f|)
2. Normalised to [0, 1] per feature
3. Fed to the LLM to produce a one-sentence alert

This gives a feature importance mask (which sensor signals drove the prediction) and a feature delta (signed change vs. 6-day baseline).

Why Not GNNExplainer?

Criterion	GNNExplainer	Gradient Attribution
Inference time	~200 ms (200-step optimisation loop per cow)	~5 ms (single backward pass)
Stability on small graphs	Can oscillate on <100-node graphs	Deterministic — identical on repeat
Feature importance ranking	Soft mask on adjacency + features	Direct ∂output/∂input per feature
Implementation	Requires torch_geometric.explain + careful hyperparams	4 lines of PyTorch (requires_grad_, backward, .grad)
Real-time API	Too slow for live GET /explain/{id}	Suitable for sub-10 ms response

For our use case — translating model output into a farmer-readable sentence — the ranking of feature importances is what matters, not the edge-subgraph mask that GNNExplainer excels at. Both methods produce equivalent top-feature rankings on our 60-cow synthetic dataset (Spearman ρ > 0.92 in our tests).

Implementation

backend/graph_utils.py  →  get_gnn_explainer_output()
backend/xai_bridge.py   →  build_explanation() + LLM alert generation
backend/main.py          →  GET /explain/{cow_id}

The function is named get_gnn_explainer_output for API compatibility with the brief's specification. Internally it uses gradient attribution as described above.

References

• Ying, Z. et al. (2019). GNNExplainer: Generating Explanations for Graph Neural Networks. NeurIPS.
• Selvaraju, R.R. et al. (2017). Grad-CAM: Visual Explanations from Deep Networks. ICCV. (Gradient attribution ancestor.)

---

Quick Start

# Backend
source venv/bin/activate
uvicorn backend.main:app --reload

# Frontend (separate terminal)
python3 app/server.py
# → http://localhost:3000

Model Training

python train.py --epochs 100 --runs 12
# Best checkpoint: models/tauron_model.pt (AUROC ~0.995)

Farmer speaks into phone/laptop mic ↓ Web Speech API (browser-native, no install) "Cow A wasn't eating much and milk was low. B gave 24 litres, fine. C is limping." ↓ POST /api/voice → Ollama/Mistral (local) ↓ Structured columns extracted:

┌────────┬──────────┬─────┬─────────────┬────────────────────────┐ │ cow_id │ yield_kg │ pen │ health_event│ notes │ ├────────┼──────────┼─────┼─────────────┼────────────────────────┤ │ A │ null │ A1 │ off_feed │ reduced milk yield │ │ B │ 24.0 │ A1 │ none │ │ │ C │ null │ A1 │ lame │ │ └────────┴──────────┴─────┴─────────────┴────────────────────────┘ ↓ Farmer reviews pre-filled rows, edits if needed ↓ Save → POST /api/ingest (one row per cow) ↓ _ingest_log → ML pipeline baseline updated