Neuralis is seeking one Ghanaian cocoa-sector partner to recruit 20–50 Asante-Twi-speaking farmers for a two-week AgriVoice pilot and name an extension officer who will own questions the system cannot safely answer. The pilot will test whether farmers receive useful, understandable responses while uncertain or sensitive questions reach a qualified person.
Imagine Kofi, a composite cocoa farmer, standing at the edge of his farm late in the afternoon. Rain looks possible. He holds an agrochemical container whose label does not settle the question in his mind: how long should workers stay out of the sprayed area?
He records the question in Asante Twi on WhatsApp. A quick answer could affect what happens on the farm the next morning. A confident but unverified answer could expose workers or damage the crop.
AgriVoice cannot safely invent the missing instruction. The responsible response is to escalate Kofi’s question to a named extension officer. But without a partner who owns that handoff, his message could sit unanswered while the decision becomes urgent.
The partner makes the safety system complete
Neuralis has built the AgriVoice code path around a deliberate boundary. Speech recognition captures the farmer’s question, then a language model selects from reviewed content blocks. It does not write agronomy advice from scratch.
That distinction matters most when the question involves pesticides, dosage, re-entry periods or pre-harvest intervals. Three chemical content blocks remain withheld until a Ghanaian agronomist verifies those details. Refusing to answer is the correct behavior while that evidence is missing, as explored in What Happens When AgriVoice Cannot Verify Safe Pesticide Guidance?.
Software can recognize uncertainty and route a question. It cannot decide who will pick up the case, contact the farmer and remain accountable for the response. That is the partner’s essential role.
The partner would recruit a defined group of farmers, name the extension officer responsible for escalations and help Neuralis understand how the workflow fits existing field relationships. This creates a real operating loop: farmer question, reviewed answer or escalation, human follow-up, measured outcome.
Twenty to fifty farmers can reveal the failures that matter
The proposed pilot is intentionally small: 20–50 farmers over two weeks. Its purpose is to produce evidence for a continue, iterate or stop decision, rather than stage a public launch.
Neuralis will measure whether at least 70% of questions are answered or correctly escalated, whether farmers understand the spoken response, whether they return in the second week and whether the exchange remains conversational. The safety threshold is stricter: no unsafe pesticide answer may reach a farmer.
The pilot will also track escalation performance. A named owner should respond in less than one working day at the median. Unresolved queues or repeated confident errors would trigger a redesign, regardless of how impressive the demonstration appeared.
That narrow scope protects farmers and gives the partner a manageable commitment. It also exposes problems that a laboratory test cannot reproduce: farm noise, code-switching, unfamiliar phrasing, unclear recordings and questions that fall between prepared topics. At least 20 real farmer questions are needed before exposure to measure speech recognition against actual field voices.
One WhatsApp message tests the whole chain
Return to Kofi at the farm boundary. His message tests more than transcription. It tests whether AgriVoice identifies a safety-sensitive question, avoids filling in a missing dosage or interval, explains that human help is needed and sends the case to someone who has agreed to receive it.
The extension officer sees the escalation while Kofi’s decision is still open. After checking the relevant guidance and the facts of the situation, the officer follows up. The next morning, Kofi has a verified instruction rather than a plausible-sounding guess.
That scene is illustrative, not a customer result. It shows the standard the pilot is designed to test.
WhatsApp access has previously supported a real end-to-end exchange, but the connection experienced a later transient restriction. A fresh inbound and outbound conversation must therefore succeed before the pilot begins. Consent, audio deletion and the full escalation route must also be rehearsed. No farmer should enter the pilot until every applicable gate passes.
The practical commitment Neuralis needs
The right partner already works with cocoa-farming communities in Ghana and can recruit 20–50 Asante-Twi-speaking farmers without treating them as anonymous test traffic. It can also appoint one extension officer who has the authority and time to receive escalations throughout the two-week period.
Neuralis will bring the voice workflow, reviewed-content controls, measurement plan and technical support. Translation must be completed by a farming-aware Asante Twi speaker, and the held chemical guidance must remain unavailable until an agronomist clears it. The partner brings the trusted human relationship that makes safe escalation possible.
This is also a supply-chain question. Programmes that aim for healthier farming households depend on advice reaching people in language they understand, with a person available when the answer cannot be automated safely. Voice access can shorten the path to information. Accountability closes the final gap.
For Kofi, success has a simple shape: he asks one urgent question in the language he uses, receives reviewed guidance when it exists and reaches a responsible person when it does not. The pilot starts when one cocoa-sector partner agrees to stand at that last, crucial handoff.
Comments
No comments yet.