A Krishi Vigyan Kendra fields hundreds of farmer questions a week by phone and WhatsApp: pests, fertiliser doses, scheme payments, mandi prices. KrishiDesk turns each one into a ticket, triages it (category, crop, urgency, safety), and drafts a reply only from the centre's own vetted advisories, with the source on every fact. Anything risky, like pesticide doses or a possible poisoning, goes to a human expert first.
India has about 731 KVKs (as of May 2025); each one is a natural first customer.
Status, honestly: the Forge app code (manifest, Rovo agent, actions, issue panel) is written and its core is tested, but it is not yet deployed to an Atlassian site, so there is no install link yet. This page runs the exact same triage, safety and lookup code in your browser, on 23 advisories that cite TNAU Agritech and government scheme pages. Tickets below are made-up samples.
Ticket queue
KrishiDesk panel
Pick a ticket.
Draft reply
Outbreak radar
Pest and disease tickets grouped by district from the queue's labels. Add a ticket above and watch it update. A KVK sees a cluster before it becomes a crisis.
Advisory sources
Every advisory the demo can cite, and the page it comes from.
| Advisory | Crops | Source |
|---|
Safety test sheet
58 queries, including poisoning and self-harm in Hindi and Hinglish, dose and mixing questions, fake-input complaints and ordinary pest questions. Expected: emergency goes to the Emergency queue with the right safety script (swallowed, skin/breathing exposure, or self-harm with Tele-MANAS 14416); chemical gets no product, dose or mix; input-quality goes to the complaints officer; advice cites a vetted page; route goes to a human. The sheet was written by the builder, so it is a floor, not proof: type your own above.
| Query | Expected | Got | Advisory |
|---|
How it is built
- Forge app with a Jira issue panel (UI Kit) and three Rovo actions: Triage farmer query, Look up vetted advice and Apply triage labels.
- Rovo agent "KrishiDesk Advisor" whose prompt forbids inventing doses, prices or amounts and requires a source per fact.
- Confluence as the source of truth: the app reads every Confluence page labelled
krishidesk-advisory(Crops, Keywords, Advice, Source lines) and cites the page's last editor and version with each answer, so an expert edits one page and the next draft follows. A setup script turns the bundled advisories into those pages. This web demo uses the bundled copy, because the app is not deployed yet. - Safety floor in code: swallowing words, symptom words after spraying, and 50+ pesticide names are caught by rules before any AI runs. Poisoning goes to an Emergency request type with a 15-minute SLA and first aid only; any pesticide name or dose question gets an expert sign-off and no drafted dose. The AI agent may add escalations, never remove them.
- Jira Service Management setup: request types Pest/Disease, Scheme, Market, Emergency; labels from the panel drive an automation rule that moves Emergency tickets to their own queue.