AgScribe

Automated information extraction from agronomist–farmer interview transcripts, powering carbon quantification and sustainability reporting.

Research paper: arXiv 2510.12023

What is AgScribe?

AgScribe is a neuro-symbolic AI system that listens to — or reads — conversations between agronomists and farmers, then automatically extracts structured agricultural data that would otherwise require manual data entry. The extracted data feeds directly into carbon quantification models and sustainability reporting platforms such as AgOutcomes.

Agronomists conduct interviews covering crop history, tillage practices, fertilizer and chemical inputs, livestock operations, and more. These conversations are rich with information but arrive in unstructured natural language. AgScribe converts them into typed, structured events — removing a significant data-entry burden and enabling faster, higher-quality carbon program enrollment.

How It Works

1. Transcription

Agronomist–farmer phone calls are transcribed into speaker-diarized text with timestamps.

2. Event Extraction

A neuro-symbolic LLM pipeline reads the transcript and identifies structured events — field data, crop decisions, practices, inputs, livestock metrics.

3. Carbon Modeling

Extracted data flows into quantification models (e.g., DayCent) to estimate greenhouse gas emissions, sequestration, and sustainability metrics per field.

4. Reporting

Farmers and agronomists receive farm reports showing conservation practices, efficiency metrics, and carbon outcomes — without manual data entry.

Operation Types Supported

Crop Farms

Row-crop operations: field boundaries, multi-year crop rotations, tillage, fertilizer & chemical programs, cover crops, irrigation, yields, and conservation practices.

Dairy Operations

Herd counts (lactating, dry, heifers, calves), milk production, protein & butterfat %, energy use (diesel, electricity), manure application, pasture management, and grazing seasons.

Pork Operations

Barn details, PQA certification, energy use (electricity, propane, natural gas), mortality disposal, renewable energy, employee count, charitable contributions, and associated row crops.

Event Types Extracted

The system identifies and structures over 25 distinct event types from natural conversation.

Field & Crop
Field (FSA/acres)
Crop Rotation
Planting
Harvest & Yield
Cover Crop
Tillage
Irrigation
Soil Testing
Fertilizer
Chemical Use
Manure Application
Conservation Practice
Livestock
Lactating Cows
Heifers
Weaned Calves
Milk Production
Protein %
Butterfat %
Grazing Season
PQA Certification
Mortality Disposal
Donations
Energy & Infrastructure
Diesel Use
Electricity Use
Propane Use
Natural Gas
Renewable Energy
Employees
Key Capabilities

Handles Natural Speech

Extracts data from messy, conversational transcripts — incomplete sentences, corrections, tangents, and filler words included.

Typed Structured Output

Each extracted item carries a precise event type, relevant fields, values, and units — ready for model ingestion with no manual parsing.

Quality Assessment

Each result includes metadata: whether expected interview sections were present, and what fraction of expected event types were active.

Multi-Species Support

A single pipeline handles crop-only farms, dairy operations, pork operations, and mixed enterprises in the same pass.

Browse a real (anonymized) agronomist–farmer interview alongside the structured data extracted by AgScribe. Select an operation type below.
Interview Transcript
Extracted Events