CCTNS — Ministry of Home Affairs

The federation problem

Most national digital projects fail at the same place: variance. India is not one design problem — it's the same problem multiplied by every variance axis you can name. Designing one product for the whole country means designing for the variance, not around it.

The trap is to design for the median user. The median doesn't exist here — variance does. Every design decision was tested against the extremes, not the middle.

Unified-with-flexibility

The strategic question wasn't "what should the citizen experience look like?" It was deeper: should there be one citizen experience, or 28?

Each state's IT department had political and operational ownership of its own CCTNS instance. The path of least resistance was to give each state a customisable citizen surface — let them re-skin, re-arrange, re-name. That would have shipped faster, with less political friction, and would have replicated the exact fragmentation that made the original system unusable.

The harder call — and the one I pushed for, at real political cost — was one unified citizen surface, with state-specific logic in the backend. One vocabulary across India. One mental model a citizen carries from Maharashtra to Karnataka to Tamil Nadu without re-learning the app.

The flexibility lives where it should — in the system's customisation slots, not its surface. States can adapt language, local terminology, minor flow steps, and integration with their existing officer systems. They cannot fork the citizen surface, the navigation vocabulary, or the data shape. The line between "you can change this" and "you cannot" was the most-debated artefact of the project.

That decision cost political capital with state IT departments who wanted their own branding on the citizen surface. Escalated, lost a battle, kept the larger war. The unified system survives.

Leadership & governance

A team of 5 designers across 5 citizen verticals (FIR, missing person, women's safety, traffic, cyber) shipping in parallel against a national-scale rollout. No two designers can build the same problem twice; no one designer can hold the whole system in their head. Governance is not a phase — it's the infrastructure.

The model that carried us:

• One designer owns one vertical end-to-end. Ownership before scale. Each designer carried a vertical from research through handoff. They knew their citizen archetype, their officer counterpart, their political stakeholders. Cross-vertical decisions came back to me, but the depth was theirs.
• Weekly system reviews. Every Friday, 90 minutes. Each designer presented one decision they were considering. We pressure-tested. Half the value was the decisions made; the other half was the team learning each other's domains.
• The system has one source of truth. Tokens, components, and patterns live in one Figma library, governed by a small architecture committee. Verticals consume — they don't fork.
• "Affects more than one team's UI" = system change. One sentence — but it's been the most-cited governance principle on the project. Anything that crosses vertical boundaries goes through the system review, not a designer's individual call.
• A short RFC process for new patterns. 1-page proposal, 48-hour review window, accept-or-reject. Speed without bottleneck. Designers shipped fast; system stayed coherent.

The hardest leadership lesson: in projects with this many stakeholders, you can't make every right decision. You make the right shape of decisions and let the team carry the details. Trusting the depth of a vertical owner is the senior move — the alternative is becoming the bottleneck for 5 verticals' worth of micro-decisions.

AI in regulated work

Designing AI-assisted workflows in any regulated domain is a contract negotiation between speed and safety. In a national policing platform, the contract is non-negotiable — get it wrong and people's lives are affected. The frame I worked from:

AI accelerates design work, not officer work. The product is what officers and citizens use. AI lives upstream — in how the team builds.

Where AI accelerated my team:

• Multilingual flow generation: drafting voice-first citizen flows in 6 languages would have taken weeks per release. I built Conversation Flow Generator — uses Anthropic's models to generate localised flow variants from a single English source. Native-speaker review still required, but the per-language draft cost dropped to hours.
• System validation at scale: with the component × language × vertical matrix we had, manual token-usage review didn't scale across 5 designers. Design System Validator flags structural breaks before engineering handoff.
• Code prototyping for kiosk + voice flows: citizen flows that needed real-time interaction (offline kiosk mode, voice playback) were prototyped in code, not Figma. CLAUDE.md Exporter packages the system as Claude Code context so prototypes generate inside our tokens.

Where I did not use AI in the product: any flow involving evidence, complainant identity, officer judgement, or case adjudication. Putting AI inside those would have been irresponsible at this scale and would have failed MHA compliance review on first pass. The boundary is a contract — speed in design tooling, deterministic UX in the product.

Trade-offs

The unified-UI decision cost political capital with state IT departments. Absorbed — escalated, lost a battle, kept the larger war. The system survives.

Voice-first flows are slower for power users (lawyers, NGOs filing repeatedly) than a dense form. Accepted. The platform's ground-level legitimacy depends on the rural farmer being able to use it, not the lawyer being optimised for.

Killing real-time chat between citizens and officers cost us a feature stakeholders had assumed would ship. Replaced with structured request/response flows that compliance could approve. The lesson: in compliance-heavy workflows, the question isn't "can we ship this?" — it's "can the regulator look at this in 6 months?"

Outcome

Pilot rollout numbers — state-level adoption, citizen usage, officer satisfaction, time-to-resolution — are part of the deeper case study available under NDA.

What I'd change

I would have invested earlier in field research with non-metro officers. We did the citizen-side ethnography well, but our officer assumptions were Delhi-coded for the first quarter, and we burned cycles redesigning officer flows once we got to ground truth in the field. The lesson stuck: in regulated multi-stakeholder work, every type of user needs first-hand contact — not just the headline one.

Also: I'd have written the governance model on day one, not month three. We had drift across the team's outputs until we sat down and articulated "what counts as a system decision vs. a vertical decision." Cost us two weeks of consolidation that should have been free.