← All work
CWMS / Civic tech, prototyped with AI
From greenlight to a live first version
Live in production. Usage data is confidential.
CWMS
Civic tech, Product, AI
My role
Design consultant, leading product logic, UX, and UI
Team
The company’s two founders, a developer, and me
Timeline
4 weeks to greenlight, 2 months to v1
Platform
Web portal and Android app
How it works
Everyday citizen features double as ward-level signals. An AI layer reads them against baseline research to set priorities and recommend actions.
My part
Shaped the product logic and signals with the founders, designed every screen for five roles, and built the first prototype with AI.
Where it is
Live, and used daily by the office, its field team, and residents.

The screens shown are from the prototype I built in Google Stitch. The production version was rebuilt in Figma. Pushkar, Ward 4 and every name and number in these screens are illustrative, used to show the full flow. They are not the client’s real location or data.
01
An elected representative’s standing rises or falls on things that rarely reach him in time: a water cut that drags on, a road nobody fixes, a promise that seems forgotten. Grievance portals collect complaints, but they can’t tell a one-off from a crisis, and complaints are only a small part of how people feel.
We built CWMS around a loop instead. Residents get useful services, and how they use them tells the office what they care about, where, and how urgently.
1
Residents use the app
Prices, jobs, news, events, and a way to report problems.
2
Use becomes signals
Read as patterns by ward and age group.
3
AI reads the signals
Against baseline research, it ranks what matters and suggests what to do.
4
The office acts
Fixes, ward broadcasts, news, and events.
5
Residents see it
Proof of the fix, and a say in whether it worked.
Trust grows, so residents use the app more
Five roles share one record:
Residents
Ward field workers
Operating team
MLA’s team
The MLA
How we sold it
We pitched CWMS to the MLA’s office as sentiment intelligence: a way to see what’s hurting residents most, act on it quickly, and show people it was handled. Knowing that this was the client’s real goal shaped every feature that followed.
02
Each feature in the citizen app had to be worth using on its own for people in a small agricultural and pilgrimage town in Rajasthan. Each one also tells the office something.
Get it fixed, by voice if they prefer
What’s going wrong, where, and how many people it affects
Be heard
What residents want next
Today’s price for their own crop
Which crops matter in which ward
Find work
Where interest in jobs concentrates, by ward and age group
A daily habit
Daily use, so the signal doesn’t come only from unhappy people
Take part in local life
Engagement and turnout by ward
Stay informed
Which updates land, and whether WhatsApp or SMS reaches a ward better
See work in progress and where money goes
Interest in development work
Say whether it worked
Whether a fix actually changed how people feel
Residents responded well, because these are things people in a town like this use every day. Signals are read as patterns by ward and age group, from people who signed up and used the app.

Daily utilities sit next to civic tools

Reporting takes a tap and a voice note

Events, with RSVPs and listen-aloud
03
The intelligence is built in layers, from what’s known before launch to what residents do every day.
Layer 1
Baseline research
Done before launch
Who lives where
Employment
How government schemes have performed
Past execution and budgets
Economic indicators
Fire stations, water tanker depots, substations
Layer 2
Live signals
From the app, by ward and age group
Feature use
Crops checked
Jobs viewed
Complaints and suggestions
RSVPs
News opens
Broadcast response by channel
Satisfaction after fixes
Layer 3
Report reliability
Keeps the noise out
Every account builds a track record: complaints made, how many a field worker verified, and how many were marked as spam. Reports from accounts with a history of false complaints only count once enough other people report the same thing.
Output
What the office sees
Every day
Priority for every issue
A recommended action, using real infrastructure
The channel most likely to reach the ward
The MLA’s daily briefing
It wasn’t meant to be right on day one
We agreed with the MLA’s team from the start that the system would need months of real use before its recommendations were reliable. It gets sharper as each ward’s history builds up.
Under the hood, it combines three things: scoring rules we designed, a commercial AI model fine-tuned on our baseline research, and the live signals fed to that model as context. The rules score each issue. The model clusters reports, explains the risk, recommends actions, and writes the briefings.
04
How the system turns a burst pipeline into the whole team’s top priority, and what happens next. Ward 4 is an illustrative example.
2
Context
The ward’s history shows water failures here before, tied to seasonal maintenance.
3
Risk
A repeat that drags on means a sharp drop in sentiment within hours.
4
Priority
It becomes the top issue for everyone on the team, ahead of older complaints.
5
First action
Tell Ward 4 it’s being fixed, and dispatch a tanker from the nearest depot.

The incident view puts the risk, the field evidence, and the recommended actions on one screen. Dispatching a tanker also sends the drafted broadcast to the ward.

A ward field worker confirms it on site, with a geofence check and a photo

Ward 4 hears it’s being fixed before they have to ask

The resident sees proof, and rates the fix. Scroll inside to see it all.
05
The team’s overview shows what’s pushing approval up and what’s dragging it down. The MLA logs in to a trimmed version: the live situation, the day’s red zones, and decisions waiting for him, with a daily brief written by the AI.

The team’s overview: priority wards, and the biggest positive and negative drivers

The MLA’s briefing: fewer menus, more decisions
06
Knowing where polish helps and where it gets in the way mattered as much as any feature.
For the MLA’s team
Dense, fast, and precise

Lots of information per screen, because the team works in it all day
Organised by stage, so nothing stalls unseen
Colour reserved for priority and status
For a farmer in Ward 4
Plain, large, and spoken

Hindi first, English one switch away
Audio on every screen, and voice notes instead of typing
Big type and one clear action per screen
07
This is the one product I’ve taken from an AI prototype all the way to production.
2
AI prototype
Built the screens in Google Stitch in about four weeks.
3
Greenlit
The prototype won the MLA’s team over, and the budget to build.
4
Design system
Rebuilt properly in Figma, as one system across web and app.
5
Two months to v1
Built with the founders and our developer.
6
Live, then refined
The detailed incident view shipped after v1.
Who built what
AI
The first prototype
Generated the first screens in Google Stitch and helped us brainstorm. That prototype won the budget to build.
Me, with the founders
The logic and the design
We decided which signals count, how reports earn credibility, and what makes an issue urgent. I designed every screen and flow across the app and the portal.
Our developer
The production build
Built the platform and implemented the logic, working with us throughout.
08
Being straight about it
CWMS is live and in daily use by the MLA’s office, its ward field team, and residents. Usage figures belong to the client, so I don’t publish them here.
Citizen app
Hindi and English, voice-first, with daily utilities, reporting, news, and events.
Live
Field verification
Ward-scoped tasks, geofence check, and photo proof.
Live
Web portal
Heatmap, incident analysis, broadcasts, news, events, and projects.
Live
AI layer
Clustering, prioritisation, recommendations, and the daily brief.
Live
What I took from it. The hardest part wasn’t any one screen. It was understanding what the client really needed, and making five kinds of user work from one shared truth, each at the level of detail they need.
The other lesson was patience. We’d told the client to expect months before the recommendations were reliable, but part of me still expected relevant results from day one. I learned the hard way that systems like this take time to give useful answers. Next time, I’d hold myself to the same timeline we set for the client.

