Inspiration
Agricultural advisers often spend too much time chasing one missing detail: finding the right programme guidance, calling the farmer, interpreting a spoken answer, and turning it into something safe to use. Farmers usually have the answer, but neither side needs another complicated portal or an untraceable phone note.
We built AcreVoice to make that small but important coordination loop easier—while keeping decisions with the people responsible for them.
What it does
AcreVoice helps Bavarian agricultural-support teams turn a consented callback into reviewable case evidence.
It starts with relevant public programme guidance, identifies the few facts still missing from a support case, and requests a callback only after consent. During the focused conversation, it captures the question, spoken answer, read-back confirmation, timestamp, and outcome.
An adviser then reviews the Evidence Passport for each field and can export confirmed information as a CSV or JSON hand-off package. AcreVoice does not decide eligibility, approve subsidies, submit applications, or replace an adviser.
How we built it
We built AcreVoice as a bilingual English/German adviser workspace using FastAPI and SQLite. The system keeps an append-only audit trail so each evidence item can be traced from its public source through the call and human review.
CALL-E powers the consented outbound callback and returns structured call results. A Strands agent, with Amazon Bedrock as its configured model runtime, normalises spoken values through field-specific tools—for example, area and yes/no responses. Deterministic safety checks ensure that only exact, confirmed answers are eligible for review and export.
The end-to-end flow is:
official guidance → consent → CALL-E conversation → Strands normalisation → evidence trail → human review → CSV/JSON hand-off
Challenges we ran into
The hardest challenge was making a voice workflow useful without making it overconfident. Spoken answers can be approximate, hedged, misunderstood, or incomplete. A polished summary alone is not sufficient evidence for a support case.
We designed the workflow to preserve the original question, spoken response, confirmation status, source, and next step together. Uncertain, refused, partial, or unconfirmed answers are held for human follow-up instead of being silently converted into a usable value.
We also had to keep the product honest about its boundaries: official guidance can suggest relevance, but it cannot be treated as an automatic eligibility decision.
Accomplishments that we're proud of
We are proud that AcreVoice is not just a generic voice bot. It has a narrow, practical unit of work: one evidence-backed farmer-support case.
Our Evidence Passport makes the hand-off genuinely reviewable, rather than asking advisers to trust a loose note. We also built a clear uncertainty route that records ambiguity without guessing, plus exportable CSV and JSON packages that can fit into existing case processes.
Most importantly, we kept a human accountable for every decision.
What we learned
We learned that trust comes from showing the limits of an agent as clearly as its capabilities. In a public-service context, the right outcome is sometimes not automation—it is a well-documented escalation to a person.
We also learned that a focused, consent-first conversation can reduce friction for farmers while creating better structured information for advisers. Designing around evidence, confirmation, and human review made the product safer and more useful.
What's next for AcreVoice: evidence for farm advisers
Next, we want to validate the workflow with agricultural advisers and expand the set of supported case questions and official programme sources. We would explore deeper integrations with existing case-management systems, stronger production security and access controls, and privacy-conscious retention policies.
The core principle will remain the same: AcreVoice coordinates the repetitive work around a case, while advisers keep the judgment and responsibility.
Built With
- agents
- ai
- amazon
- bedrock
- call-e
- fastapi
- javascript
- python
- sqlite
- strands
- voice
Log in or sign up for Devpost to join the conversation.