The Deployment
Mahavistaar is an agricultural advisory service for farmers in Maharashtra. In January 2026, we deployed VoicERA as the backend for their voice helpline — replacing a human-staffed operation that could only function during business hours.
The pilot ran for 6 weeks. Here is what we found.
The Numbers
- Call connection rate: 88.2%
- Average call duration: 2.8 minutes
- Languages handled: Marathi (primary), Hindi, English
- Peak concurrent calls: 47 (out of 50 capacity)
- Human agent escalations: 3.1% of calls
- Uptime: 99.94%
The 11.8% call failure rate was primarily network-level drops (poor rural connectivity), not system failures.
What Farmers Actually Asked
The top query categories, in order:
1. Pest identification and treatment
2. Weather-based spray timing
3. Government scheme eligibility
4. Seed variety selection
5. Market price information
Query category 4 and 5 required knowledge base updates after launch — we hadn't anticipated the breadth of scheme and pricing questions.
What Surprised Us
The conversations were longer than expected. We designed for 90-second queries. The average was 2.8 minutes. Farmers did not just ask one question and hang up. They followed up, clarified, asked related questions. The multi-turn conversation capability was more used than we anticipated.
Code-switching was constant. Marathi sentences frequently included Hindi words. Our IndicWav2Vec model handled this reasonably, but there were failure cases that pure Marathi training data would not have prepared us for.
The preference for voice over text, even when both were available. When we offered SMS follow-ups with written summaries, only 12% of callers opted in. The farmers trusted and preferred the voice interaction.
What We Changed After Week 2
- Expanded the knowledge base to cover 340 government agricultural schemes
- Added a confidence threshold — queries below 70% confidence now route to a shorter "I'll connect you to an expert" response rather than a low-quality answer
- Improved Marathi-Hindi code-switching recognition with a custom acoustic model fine-tune
The Core Finding
A phone call, in a farmer's own language and dialect, that answers their question within 2 seconds — is genuinely useful. Not "good for AI". Not "impressive for a demo". Actually useful, in a way that changes farming decisions.
That is what we set out to build. The Mahavistaar pilot confirmed we had built it.