Mend | Devpost

Inspiration

Every day on our way around USC we step over cracked sidewalks and swerve around the same potholes. We wondered why damage everyone can see stays unfixed for months. So we asked. In a survey of 51 people, 92% had noticed a street problem in the past month, but only 1 in 10 reported it through an official channel, and 76% did nothing at all.

What it does

Mend turns one photo into a verified repair report.

Snap: take one photo of the damage. There are no forms.
Auto-fill: an on-phone AI model identifies the type of damage, and GPS plus the compass pin down the exact spot and the road segment or asset you're pointing at.
Verify: a server-side vision model confirms the report and scores its severity from 1 to 5. A repeat report of the same problem counts as a confirmation, not a new ticket.
See it fixed: when the city marks the problem fixed, you get a notification.

Reporters earn points they can spend at local cafés and shops. The first person to report a problem earns full points; later reports earn a few for confirming it's still there. On the other side, a web dashboard gives cities and business districts a work queue ranked by severity. Residents use Mend for free. Stores pay about $2 per visit when someone redeems points with them, and business districts and city departments pay for clean, de-duplicated data.

How we built it

iOS app in SwiftUI, with a live Supabase backend for the database, storage and server functions.
Two on-device damage models (a classifier and a detector), trained on about 38,000 public images.
On-device speech-to-text for optional voice notes. Reports taken without signal are queued and sent later.
A server-side vision model that gives the final verdict on damage type and severity.
A web dashboard where buyers can mark reports fixed and post bounties to boost coverage in chosen areas.
Research: a survey (51 responses, Sep 25–27), an analysis of 1.77M LA 311 requests (Jan–Sep 2026), and a scan of 311 apps, 311 software vendors, road-scanning tools and rewards apps.

We verified the full flow, from photo to AI verdict to points, in the simulator, with demo reports seeded around USC.

Challenges we ran into

Our survey overturned our assumptions. We expected game features to drive usage, but only 4% of respondents named them, and only 24% were motivated by rewards. We rebuilt the pitch around speed and seeing fixes.
Designing rewards that don't reward spam. Paying for every report would have encouraged duplicates and fake reports. We settled on: full points only for the first verified report, in-app camera only, GPS and a timestamp, a daily cap, and points held until a report is verified.
Messy city data. In the 311 data, a quarter of illegal-dumping reports closed with nothing found and about a third of graffiti reports were duplicates. Understanding why shaped our verification design.

Accomplishments that we're proud of

A working app with a live backend.
AI running on the phone, so reporters see a suggestion instantly, even with a weak signal.
Grounding every feature in evidence: each step of the app maps to a barrier from our survey (73% said reporting takes too long, 57% didn't know who to report to, 43% believed nothing would be done).
A business model where residents never pay, and where the rewards rules produce cleaner data for the city.

What we learned

People care more about their neighborhood than about prizes. 63% are motivated by safer places they visit; rewards matter far less.
What people say they'll do isn't what they do. 76% said they'd try the app, but only 10% report issues today. Closing that gap is the real test.
Speed is the product. 90% will spend two minutes or less on a report, so every extra field loses people.
A city isn't the only customer. Local stores and business districts benefit when problems get fixed, and they can move faster than a city contract.

What's next for Mend

A two-week pilot near USC with 100 students, half with points and half without, counting actual reports rather than stated intent.
A pay-per-visit trial with 5–10 cafés near campus to measure how often points get redeemed and what cafés will actually pay.
Sending pilot reports through 311 to test whether they come back with fewer duplicates and "nothing found" closures.