CivicPulse AI — Interactive Prototype

9:41

Ward 12 · KothrudPune Municipal Corporation

Ward health score

72/ 100

↗ ▲ 6 pts vs last month

41 of 60 closures were verified

Verified resolution rate 68% · ward target 80%

Avg. verification time 6.2 hrs

My complaints 4 OPEN

Pothole · Paud Road

CPL-2026-04871 · raised 6 days ago

AI verificationPWD

92%
conf.

Street light out · Lane 7

CPL-2026-04720 · raised 11 days ago

Awaiting your check

Tap →

Water leak · Sutar Chawl

CPL-2026-04355 · closed 2 days ago

✓ Verified resolution

100%

Nearby issues WITHIN 1 KM

Garbage · 300 m

17 reported by 17 people

Open drain · 800 m

SLA 2d officer assigned

What you're looking at

A guide to every element on the screen to the left.

Citizen · Home

Ward health score

The citizen-facing roll-up of their ward: how much of what was closed was actually verified. It's the same north-star number the ward admin is measured on, so residents and officials read one number, not two.

41 of 60 verified

Deliberately shown as a fraction, not a percentage alone. A closure without evidence is visible as a gap, which is the whole product thesis.

Verification prompt

Appears only when a complaint has passed AI verification but is still waiting on the citizen. Their confirmation is a required gate — nothing closes silently behind their back.

My complaints

Each row carries the live workflow stage as a chip, plus the AI confidence score where one exists, so the citizen sees exactly which of the eight timeline states they're in.

Nearby issues

Surfaces open issues within 1 km. Repeat-report counts feed duplicate detection: reporting something already logged joins the existing complaint instead of creating a new one.

Citizen · Report

Live viewfinder only

This is the in-app camera, not a file picker. The gallery path does not exist anywhere in the flow — it's the single design decision the anti-fraud model rests on.

Burned-in watermark

Coordinates, timestamp, heading and IDs are drawn into the frame at capture and also signed as metadata. A screenshot of this image is detectable because it carries no camera-sensor signature.

Attached metadata

The seven fields the spec requires on every capture — lat, long, timestamp, accuracy radius, heading, ward ID and complaint ID — shown to the citizen so the record is not a black box.

AI extracted

Category, severity, department and ward are inferred from the photo plus the voice or text note, so the citizen never picks from a dropdown of municipal department names.

Duplicate check

Runs before submission against open complaints in the same geographic cluster. A match would offer to join that complaint rather than open a second ticket.

Citizen · Track

AI confidence

The verification engine's composite score. Above 85% it auto-passes to citizen verification, 60–85% goes to the ward admin's manual review queue, below 60% is rejected back to the officer.

Check-by-check list

The score is never a bare number. Each contributing check — GPS, timestamp, EXIF integrity, object detection — is shown individually so a dispute has something concrete to point at.

Before / after pair

The officer's own captures, side by side. Scene-change comparison between these two frames is what distinguishes a real repair from a photo taken at a different spot.

Eight-state timeline

Submitted → Assigned → Accepted → Started → AI verification → Citizen verification → Verified → Closed. The two verification states are the ones this product adds; everything else exists in any complaint system.

Citizen · Verify

The last gate

AI clearing the evidence is not closure. Where a citizen check applies, their confirmation is required, which is what makes "verified resolution" a claim about reality rather than about paperwork.

Officer submission, shown first

The citizen sees exactly what was submitted on their behalf before being asked to agree with it.

Fresh capture required

Same rules as reporting: live camera, at the location, no gallery. A verification photo taken from home would fail the GPS check.

Three-way comparison

Original complaint photo, officer's after photo, and citizen's verification photo are compared pairwise. Agreement across all three is much harder to fake than any single image.

"Still broken" reopens

A rejection here reopens the complaint and counts against the officer's first-time verification rate — the mechanism that makes premature closure expensive.

Officer · Tasks

Four-number dashboard

Assigned, pending verification, high priority, completed today. "Pending verification" is its own count because work submitted is not work finished under this model.

Rejection banner

AI rejections come back to the officer as required work, not as a notification they can dismiss. The complaint cannot progress until it is recaptured.

Route ordered by SLA

Sorted by time remaining rather than distance, with distance shown so the officer can batch nearby stops themselves.

Row-level signals

"In geo-fence" tells the officer capture is already unlocked at their current position. "Recurring 3rd time" warns that a patch fix here will likely bounce back.

Officer · Capture

Geo-fence gate

The red pin is the complaint, the blue dot is the officer, the dashed ring is the configured radius. Outside it, the capture buttons are disabled — the camera physically will not open for this complaint.

Before and after slots

Both are captured live on site. The before photo is taken when repair starts, not retrospectively, which is what makes the pair meaningful as evidence.

Pre-flight checks

GPS lock, server-synced timestamp, a second viewing angle, and image quality. Quality and angle checks run at capture so failures surface on site, while the officer can still fix them.

Completion video

Optional or mandatory per category. Roadwork requires it; a street-light replacement would not.

Submit stays disabled

Until every mandatory check passes. There is no "submit anyway" path — incomplete evidence never enters the verification queue.

Officer · Result

Rejected, with a reason

The engine returns one of three outcomes — Verified, Needs Review, Rejected. A rejection always names the specific failing check rather than returning a bare score.

Object still detected

The strongest single signal: the same object class is present in the same frame region of the after photo. This is the check that catches "photographed a different stretch of road".

Scene change 8%

Before and after are nearly identical. Genuine repairs produce a large, localised change in exactly the region the complaint was about.

Passing checks still shown

GPS and EXIF passed. Showing them keeps the rejection specific and reviewable, and gives the officer a real basis for the dispute path.

Cannot resubmit the same frames

Image hashes of rejected evidence are retained, so re-uploading them trips duplicate detection. The only way forward is a revisit.

Ward Admin · Overview

Verified resolution rate

The north-star metric, given the whole hero card. It is deliberately lower than any "complaints closed" number the ward has ever reported — that gap is the point of the product.

146 of 214

The denominator is closures, not complaints, so the metric can't be improved by leaving tickets open. The 68 unverified closures are named explicitly.

Supporting metrics

First-time verification rate, average verification time, citizen satisfaction and reopening rate. Together they distinguish a ward that resolves well from one that merely resolves fast.

Recurring issues

Locations that get verified-fixed and then break again. A high verified rate alongside a high recurrence rate means the repair standard is wrong, not the officers.

Budget recommendation

Derived from recurrence cost — the accumulated spend on repeat patch repairs against the cost of a permanent fix. This is where verified history turns into a procurement argument.

Ward Admin · Map

Heat map

Coloured by unverified backlog rather than raw complaint volume, so a busy zone that clears its work reads cool and a quiet zone that closes without evidence reads hot.

Hot zones

Each zone carries its own verified rate. Market Rd at 41% is the actionable one: high volume, past SLA, and closures that aren't holding up.

Officer utilisation

Load per officer, red above 90%. Sustained overload is the leading indicator of rushed evidence and AI rejections downstream.

Rebalance

Reassigns open complaints across the ward using load, geography and each officer's verification success rate.

Ward Admin · Alerts

Fraud detection

Duplicate image hashes, screenshot detection and metadata mismatches, each tied to a complaint and an officer. These are the cases the automated checks caught but a human must decide on.

Duplicate image

The same after-photo submitted for two different complaints. Perceptual hashing catches this even after crops, filters or recompression.

Screenshot detected

The frame has no camera-sensor signature, so it was displayed on a screen and re-photographed or exported. Liveness detection fails it before it reaches verification.

Needs review queue

The 60–85% confidence band. AI won't pass or reject these on its own, so a human breaks the tie — and the decision becomes training signal.

Escalation

SLA breaches and unresolvable disputes move up to the zonal office with the full evidence chain attached.

Prototype only — all data is illustrative. Source: prototype/index.html, single self-contained file, no dependencies.