Digital Green · FarmerChat

A farmer gets advice.
Did it help?

It's a simple question, and a hard one to answer for millions of people. Field Echo asks it gently, one farmer at a time, about the advice they were actually given.

  • Vaishnavi Managoli
  • Ishanth Thota
  • Dhanshree Atre
  • Daniel Epstein
  • Benjamin Young

Built in 2 days at the AWS Breaking Barriers hackathon

The problem

Giving advice is the easy part.
Knowing if it helped is not.

FarmerChat answers millions of farmers in their own language. But understanding whether that advice was followed, and whether it worked, still depends on field studies that take months and only ever reach a small sample.

0
farmers rely on FarmerChat

Studies take months

A single field study runs for months and reaches only a few thousand people.

One survey for all

Everyone gets the same questions, no matter what they actually asked about.

So most farmers are never asked, and the ones who are rarely get a question that fits.

Our idea

A follow-up that fits
the advice each farmer received.

  1. 01

    FarmerChat remembers each conversation, the question the farmer asked, the advice they were given, and how it was resolved.

  2. 02

    From that, the agent writes a short survey about their advice, in their language, at a point in the season when the answer will mean something.

  3. 03

    Each survey is checked before it's sent. If a conversation has nothing worth following up on, the farmer simply isn't contacted.

A good question at the right moment, and nothing more.

FIELD ECHO · MEL SURVEY AGENT Deployment on AWS Digital Green · FarmerChat Team 37 · AWS Breaking Barriers 2026 AWS Cloud us-east-1 Default VPC Private: only CloudFront reaches the server Amazon EC2 · t3.small instance fieldecho · Docker Compose HOST SERVICE nginx reverse proxy CONTAINER MEL Console CONTAINER PostgreSQL 16 runs, configs, scheduler outbox Survey agent · production Decide golden rules Generate 1–3 questions Check survey JSON Review reviewer agent Hand off save, schedule at most one redraft Pydantic AI · in the console · saves to PostgreSQL Console users team and Digital Green in a web browser HTTPS CloudFront fieldecho.org HTTPS, ACM cert VPC origin private sign-in Amazon Cognito fieldecho user pool Amazon Bedrock OpenAI · Anthropic · Amazon · Google models model calls AgentCore Runtime AWS MANAGED Survey agent · Strands build compares models · Agents tab logs in CloudWatch GitHub Actions main branch only 1 · push exact-SHA image IAM OIDC role 3 · send the fixed deploy command Amazon ECR dg37-explorer immutable tags AWS Systems Manager Run Command DG37-DeployMain-v1 pull image 4 · deploy, roll back on failure 2 · agent code release Amazon S3 CodeZip update runtime Checked against the AWS account on 5 Nov 2026
The rules that decide

Before any survey is written,
plain rules check the insight.

Most of these end the process early, so the common case costs nothing. Only the last one leads to a survey.

R1
Junk or invalid insightBroken or low-quality data is dropped.
R3
Nothing to verifyGeneral information with no action to follow up.
R4
Waiting for detailsThe conversation isn't finished yet.
R5
No check neededThe advice doesn't call for a follow-up.
R6
Season has passedNo sensible send window remains before season's end.
R7
Crop already harvestedThe calendar shows the crop is done.
R8
Send a surveyNone of the above applied, so a survey is worth writing.
What it costs

The same $12 buys
one phone call, or 200 follow-ups.

6¢
per survey, written by one model and checked by another.
0
farmers Field Echo can follow up with for $12.
0
farmer reached for that same $12 by a human phone interview.

Cheap enough to ask the whole base, not just a sample.

Human cost: Innovations for Poverty Action (IPA), Phone Surveying in a Pandemic Handbook, 2020. AI cost measured from our own run.

Built in 2 days at the AWS Breaking Barriers hackathon, November 2026. Open source under the MIT License: github.com/AWS-DG37/AWS-breaking-barriers-DG37