CPGRAMS

Most products begin by inventing a reason to exist. This one did not. CPGRAMS is already a constitutional-grade commitment: any citizen of India can lodge a grievance against any central government department, and an officer is obliged to answer it, usually inside 30 to 60 days, with automatic escalation to senior officers and a route to the Prime Minister's Office if they do not.

It is not a pilot or a portal somebody is trying to get adopted. It runs across more than 90 ministries and handles over 20 lakh grievances a year, and it disposes of 93% of them. The machinery works.

So this project was never about designing a service. The service exists. It was about the fact that the door into it could only be opened by people who least needed it.

To use that machinery, a citizen has to fill in a form. One page, fifteen or more fields, written in departmental language, on a layout built for a desktop computer.

The hardest field is the second one. Before describing anything, you must name the ministry and the category your problem belongs to. That is a filing decision. Ask a person whose pension has stopped which of ninety departments owns that, and the conversation is already over.

Language is the clearest exclusion but not the only one. A quarter of the country cannot read or write at all, which makes any text interface a closed door regardless of which language it is written in. The 60-plus age group files the most grievances and has the lowest digital literacy, so their complaints get filed by somebody else, or filtered, or delayed, or never made.

Put those together and 78% of grievances arrive from urban, educated users.

That is not a usability score. It is a democratic problem. The feedback loop between a government and its citizens was quietly sampling only the citizens who were already doing fine.

That reframe is the entire project. It moves the work from redesigning the portal to building a translation layer over it.

Nothing about the government changes. Same ministries, same categories, same statutory clock. What changes is who is required to understand any of it. The citizen describes what happened to them, in whatever language they think in, out loud if they cannot write. The system does the filing.

A chat window is still an interface, and to somebody who has never used one it is still an exam. So the product has a face.

Samadhan Didi is a government worker in a saree with a departmental lanyard, and every part of that is a decision. Didi means elder sister. She is the person you already ask for help with a form, the one at the counter who does not make you feel stupid for asking. She is lip-synced to the spoken reply, so the answer is watched as well as heard, which matters when the person listening may not be able to read the same words on screen.

She is not an ornament on the product. She is the onboarding.

This is the flow a citizen sees once, the first time they ever open the chatbot, and it carries more weight than anything else in the product. Everything after it assumes the person knows they can press a button and speak. Nothing in their experience of government websites has ever suggested that.

So the tutorial does not describe the interface. It points at it. The screen dims except the one control being discussed, Samadhan Didi stands beside it and says what it does in plain language, and the whole thing can be skipped from the first frame by anyone who does not need it.

For a quarter of the country, reading and writing is the barrier. Voice is not a convenience feature layered on top of this product. It is the accessibility strategy, and the typed flow is the alternative rather than the default.

Press once, speak, and the system does the rest. Speech-to-text runs through Bhashini across all 22 scheduled languages, the language is detected rather than selected, and the reply comes back spoken as well as written. Filing a grievance becomes about as difficult as making a phone call, which for this audience is the correct level of difficulty.

Voice is the priority, not the requirement. Plenty of citizens can type and would rather, and speaking a complaint out loud is not always possible in a shared house, an office, or a queue.

The typed path reaches the same destination through the same architecture. Describe it in plain language, let the system infer the classification, review before submitting. It is a conversation, never a form, and the difference from the portal is that the burden of knowing how government is organised never moves onto the person typing.

A component set built for a conversation rather than a page: message bubbles by speaker, audio players with waveforms, state and language tags, the review card, tutorial spotlights, and the mascot in each of her states.

It is live at cpgramsaichatbot.com, running against the real grievance system rather than standing as a concept.

For a citizen, the qualification required to complain to their own government dropped from reading English or Hindi and knowing how ninety ministries are organised, to being able to speak. For DARPG, grievances now arrive pre-categorised and correctly routed, which is work that previously landed on an officer before the statutory clock had even started.

The work was scoped against projections of 3x more grievances filed, a 40% reduction in incomplete submissions and 85% satisfaction. Those are targets rather than results, and post-launch numbers sit with the department.

Designing for government taught me that accessibility is not a layer applied to a finished product. Here it was the product. Remove voice and you have not built a slightly less inclusive chatbot, you have rebuilt the thing that was already failing.

It also changed how I think about automation. The instinct with a form this painful is to remove it completely and let the system handle everything silently. But the moment software files a legal document on somebody's behalf, hiding its reasoning stops being convenience and becomes exposure. The review screen is the least clever thing in this project and almost certainly the most important.

And working at national scale reframed what a design decision costs. A dropdown that confuses 5% of users is a usability issue in most products. On a system serving 1.4 billion people it is tens of millions of citizens who never get heard, which is a different kind of number to be responsible for.