YIELDWISE

THE COLAB IS : https://colab.research.google.com/drive/1RNsokNnPXZLm_buU1pM3klDmY2-IqQmY?usp=sharing YIELDWISE is a field-planning workspace that brings a grower’s field record, seven-day weather, transparent operating assumptions, and measured research models into one interface. It is built to make the source and limits of every result clear.

For judges: start on the Overview, open Yield outlook, then use Harvest planning and Field conditions. The Data & methodology page explains exactly which results come from research models and which are user-entered calculations.

What it demonstrates

• Field-first workflow: save a location, crop, variety, area, and field management details directly in the browser.
• Live weather context: retrieve seven-day Open-Meteo weather for the saved coordinates, with clear availability and cache status.
• Research-model access: inspect registered maize and SoyNAM targets, their held-out errors, and their supported research scope before submitting a research input form.
• Transparent planning: compare harvest-date operating returns from the quantity, saleable fraction, price, and cost assumptions entered by the user; check measured grain factors against buyer thresholds.
• Responsible product design: unavailable data stays unavailable. The app does not replace it with synthetic forecasts or present research-only models as validated farm predictions.

Quick start

YIELDWISE uses a React/Vite frontend and a local FastAPI service. You need Node compatible with Vite 8 and Python 3.11 or newer.

npm ci
python -m venv .venv
.venv\Scripts\python -m pip install -r backend/requirements.txt
.venv\Scripts\python backend/serve.py

In a second terminal:

npm run dev

Open the local URL printed by Vite. The frontend sends /api requests to http://127.0.0.1:8000; the backend binds only to your machine’s loopback address.

Judging guide

Area	What to look for
Overview	Saved field identity, weather status, and model-service readiness in one place.
Yield outlook	Target availability, held-out error, research scope, and input validation before a prediction is requested.
Harvest planning	Explicit, editable return calculations; changing a date does not invent a yield, grade, or price change.
Field conditions	Weather-based freezing, heat, and rainfall threshold checks that are labeled as checks, not crop-damage probabilities.
Data & methodology	Model coverage, scope, limitations, and data-handling details.

Technical approach

Layer	Implementation
Frontend	React 19, Vite 8, React Router, and custom CSS
API	FastAPI with Pydantic request validation
Model serving	Verified JSON artifacts loaded through a registry; no executable pickle files
Weather	Open-Meteo retrieved server-side with a bounded ten-minute cache
Field persistence	Browser localStorage; no account or cloud database required
Deployment	Docker can serve the built frontend and API on one origin

The included model artifacts are already available in backend/artifacts/. The importer validates an export’s SHA256, protocol, schema, source hashes, and saved inference cases before installing an update:

python -m backend.import_artifacts PATH_TO_EXPORT.zip --sha256 SHA256_FROM_COLAB

See the technology stack for Docker instructions and the model review for the training and evaluation record.

Research status and limits

The current registry includes five maize targets and two SoyNAM targets from measured trial data. The verified research artifacts replay 35 selected cases locally. These results are research-only: the selected maize yield model did not beat simple held-out baselines, soybean yield improved only slightly over its baselines across seven test environments, and the selected soybean R8 maturity model did not beat its strongest baseline.

YIELDWISE does not claim validated next-season yield forecasts for arbitrary farms. It also does not calculate premium-grade probabilities, optimal harvest windows, causal weather or irrigation effects, crop prices, or trading signals. See the full integration notes for the evidence and remaining validation requirements.

Field profiles stay in browser storage. Coordinates are sent to the YIELDWISE backend to retrieve weather and, when selected by the user, to OpenStreetMap. NASA imagery, NDVI, surveyed terrain, and laboratory soil data are not connected.

Checks

npm test
npm run lint
npm run build
python -m unittest backend.test_inference backend.test_api backend.test_real_artifacts backend.test_weather
python -m unittest discover -s model_training -p 'test_*.py'

The real-artifact tests skip when artifacts are intentionally removed. Training sources and the recovery-packet builder live in model_training/; prepared datasets and checkpoints are excluded from the repository.