योजना दीदी
Yojana Didi
A Hindi-first scheme copilot for Uttar Pradesh: eligibility in one conversation, documents and the exact counter, and a live lookup on official portals when the answer changed this week.
- Languages
- Hindi · English
- Channels
- Inbound and outbound phone · Missed call · WhatsApp · Browser voice call
The problem
Uttar Pradesh runs Central and State welfare schemes across eDistrict, SSPY, the PM portals and individual line departments. The schemes are not the constraint. The journey from being eligible to actually receiving the benefit is.
Three gaps repeat. Households do not know what they qualify for, because scheme information reaches them informally or not at all. Applications start and stall on pending eKYC, a missing Khatauni or the wrong counter. And nobody upstream can see where the stall happened, because status lives in portals citizens cannot navigate and adoption data reaches decision-makers a quarter late.
All three are reading problems dressed up as process problems. The household that needs the scheme most is the one least able to work a portal.
What AiSewak built
योजना दीदी is a voice agent that runs the eligibility conversation instead of the form. The citizen speaks; she asks for an age band, occupation, district and family situation, matches that profile against scheme rules, and names what the household is likely to qualify for.
She then does the part a brochure cannot: which documents, and which counter — eDistrict, a Jan Seva Kendra, the tehsil or the bank. Eight flagship schemes are covered in the public demo, including PM-KISAN, Ayushman Bharat PM-JAY, PMAY urban and rural, the UP Chief Minister's housing scheme, PMFBY crop insurance, SSPY social pensions and eDistrict certificates.
The agent is deliberately voice-only on this page. The widget runs an audio orb rather than a video avatar, because a helpline on a low-bandwidth phone should set the expectation of a phone call, not a video call.
Every conversation writes structured data — who asked, what matched, what is pending — which is what makes a district-level adoption view possible at all. A companion command-centre view sits at /yojana-didi/dashboard; the figures shown on it are labelled illustrative pilot-design figures, not measured results.
How it runs
Conversational layer
An ElevenLabs Conversational AI agent, Hindi-first with a language lock: it asks once which language the citizen wants and stays in it for the whole call rather than drifting between Hindi and English mid-sentence.
Live official-source lookup
When a caller asks about a changed amount, a new date or the latest announcement, the agent calls a server-side search tool restricted to an allowlist of government domains — up.gov.in, edistrict.up.gov.in, sspy-up.gov.in and the central scheme portals. The prompt forbids speaking a number that neither the knowledge base nor that tool supplied.
Guardrails
The agent discloses that it is automated on every call and never presents itself as an officer, a minister or the Chief Minister. It never asks for a full Aadhaar number, a bank account, an OTP or a password, and collects only a name, district and callback number with consent. Approval, amount and timeline are always deferred to the department.
Compliance review
Conversations are scored automatically for compliance across the whole set rather than a sample, and transcripts plus structured outcomes are retained in an auditable form.
What a department can replicate
- 1
The pilot shape published on the demo page is two districts — one urban, one rural — carrying eight flagship schemes with voice and WhatsApp flows and deep links into eDistrict and SSPY, so the knowledge base is validated against real citizen questions before it is scaled.
- 2
The second phase standardises the scheme knowledge-base template so a new scheme is a content change rather than a rebuild, adds a call-centre copilot for officers already on the phone, and opens district dashboards to District Magistrates and Chief Development Officers.
- 3
For a department starting from scratch, the reusable part is the pattern, not the content: a scheme-agnostic eligibility conversation, an allowlisted live-lookup tool so the agent is never the reason stale information is spoken, and a refusal rule that sends approval questions back to the department.
Scope the same pilot for your department
Two districts, a bounded knowledge base, a written list of what the agent will refuse to do, and a measurement step in the call itself. That is the shape every pilot above started from.