सूरज
Suraj
An outbound Hindi awareness agent for PM Surya Ghar Muft Bijli Yojana that explains the subsidy slabs accurately, corrects the solar-is-free misconception, and logs interest for a human follow-up team.
- Languages
- Hindi
- Channels
- Outbound phone · Browser voice call
The problem
PM Surya Ghar Muft Bijli Yojana pays a substantial rooftop solar subsidy directly into a household's bank account, and most households eligible for it have never had it explained to them in a form they can act on.
The scheme also carries a specific misconception that damages it: that solar is free. It is not. A household that installs on that belief and then discovers its share of the cost is a household that tells its neighbours the scheme is a trick.
Awareness work of this kind is outbound by nature, which makes it the exact category of calling that TRAI regulates and that citizens are right to be suspicious of. An outreach agent that does not handle disclosure, calling hours and opt-out correctly is a liability regardless of how well it explains the scheme.
What AiSewak built
सूरज is an outbound Hindi awareness agent, written to sound like a knowledgeable neighbour explaining something useful rather than a government announcement being read out. It mirrors the citizen's energy: conversational with a talkative caller, brief with a terse one, and a respectful exit with a busy one.
The scheme facts it works from are fixed and bounded, and the prompt forbids exceeding them: thirty thousand rupees per kilowatt for the first two kilowatts and eighteen thousand for the third, so up to seventy-eight thousand rupees for a three kilowatt system, paid into the bank account; up to three hundred free units a month, which takes most household bills close to zero, with surplus sold back to the DISCOM; a collateral-free loan at around seven per cent through Jan Samarth or a bank; registration on pmsuryaghar.gov.in, a registered vendor, installation, net meter and inspection, and subsidy typically fifteen to thirty days after commissioning.
The honesty guardrails are part of the scheme's own official communication rather than an afterthought. The not-fully-free clarification is built into the prompt as a misconception guardrail, and the agent is instructed never to inflate a figure.
The agent does not attempt to complete anything. It captures citizen interest through a dedicated tool for the human follow-up team, and honours an opt-out immediately through a do-not-call tool, so a refusal is recorded rather than re-dialled.
How it runs
Conversational layer
A Cartesia Line SDK agent, distinct from the ElevenLabs stack used by the scheme copilots. Speech is synthesised with sonic-3 on a private cloned Hindi voice supplied by the operator.
Deployment
The agent is a single checked-in Python entry point deployed by a repeatable script onto a separate Cartesia workspace, with the browser demo minting workspace-scoped tokens from its own endpoint. Keeping the account boundary explicit in code is what stops a demo drawing on the wrong workspace's voice and quota.
Tools
Two, deliberately: one that records a citizen's interest for follow-up, and one that marks a number do-not-call. There is no booking, no payment and no application submission — the application happens only on pmsuryaghar.gov.in.
Compliance
The agent promotes a government scheme but is not a government officer; it discloses that it is an AI assistant, never asks for money or an OTP, and campaigns are constrained to the TRAI calling window of 10 AM to 9 PM. The public page is a technology demonstration and uses no real citizen data.
What a department can replicate
- 1
For a scheme-awareness campaign, the transferable structure is a bounded fact set plus an explicit misconception guardrail. Most outreach failures are not hallucinations; they are an over-enthusiastic but technically true framing that the citizen later experiences as a lie.
- 2
Second, keep the agent's tool surface small. An outreach agent that can only note interest and honour an opt-out cannot commit the department to anything, which is precisely what makes it deployable without a legal review of every transcript.
- 3
Third, treat the calling window, the disclosure and the do-not-call list as build-time constraints. A department can then evaluate the agent on whether it explains the scheme well, rather than on whether it is going to cause a regulatory problem.
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.