Inspiration
Is there a pothole everyone on your street knows to dodge, or a puddle you keep stepping into on your way to lectures? Everyone notices it, everyone talks about it, yet it becomes part of the daily routine.
We realised that sharing a place doesn’t always mean sharing responsibility for it. A student, commuter and resident might encounter the same problem without their observations ever coming together.
LocalLoop grew from a simple idea: we already share the same places—what if we could care for them together? We wanted to give people a way to understand their neighbourhood, contribute what they know and collaborate with their local council.
What it does
LocalLoop is a community map that turns separate observations into shared action.
Residents can photograph a local issue, describe it and mark its location. Before publication, the app removes embedded photo metadata and automatically blurs detected faces and number plates. An optional AI Scan suggests an editable title, description, category and severity estimate.
Other residents independently check whether the issue exists. On-site checks carry more weight, older evidence gradually loses weight, and users cannot verify their own reports. Similar reports can be consolidated, while recurring problems and clusters of distinct issues help reveal patterns.
Qualifying reports feed into evidence-based drafts for the responsible council department. A moderator reviews the draft and downloads a PDF or supporting photo package for manual submission.
A claimed repair does not immediately close the issue: independent residents must check the outcome first.
A separate news layer adds local context. Our prototype demonstrates this with labelled sample stories; AI-assisted mapping of real news events is a planned next step.
How we built it
We built LocalLoop as a full-stack web application using Next.js, React and TypeScript, with Tailwind CSS for the responsive interface.
The interactive map uses MapLibre, a CARTO basemap and OpenStreetMap data. Server routes handle reporting, voting, duplicate detection and status changes, keeping important rules out of the browser.
For photo privacy, we combined ONNX Runtime, an UltraFace face detector, an RT-DETR number-plate detector and Sharp for image processing. Detection runs locally on the server. If it fails, the app blurs the whole image before publishing.
We integrated Gemini for optional photo interpretation and council-draft writing. Photo analysis uses the already processed image, and people can review the suggestions. Council drafts use report records as their factual input, with a template fallback when AI is unavailable.
The prototype uses demo accounts and a local JSON store, allowing us to concentrate on building and demonstrating the complete community workflow.
Challenges we ran into
One challenge was deciding what “confirmed” and “fixed” should mean. A single report should not automatically establish a fact, and a repair claim should not automatically establish success. We built separate verification stages, independence rules and visible score breakdowns to make those distinctions clear.
Photo privacy was another challenge. Detection boxes can overlap, models can miss details, and inference can fail. We added padded detection regions, overlap merging and a whole-image blur fallback while keeping processing free of paid API requirements.
We also had to distinguish repeated reports of one problem from several different problems nearby. Those cases look similar on a map but should contribute differently to community evidence and escalation.
Finally, we had to keep AI useful without giving it authority over the workflow. It can suggest and draft; people and explicit application rules determine what happens next.
Accomplishments that we're proud of
We’re proud of building a complete loop rather than stopping at a map of complaints: report → independently check → respond → independently check again.
We implemented local face and number-plate blurring without requiring a paid detection service, and made processed photographs available in a full-image viewer.
We also connected individual reports to a practical council handoff, including department-specific drafts and downloadable evidence packages.
Most importantly, different people can contribute to one outcome without knowing each other first. Someone notices, someone confirms, and someone checks the repair. Their shared location becomes a reason to work together.
What we learned
We learned that community engagement needs more than a place to post. People need clear ways to contribute, understandable evidence and a way to follow what happened afterwards.
We also learned that more reports do not necessarily mean more independent evidence. Duplicate posts, separate incidents, recurring failures and fresh observations all need different treatment.
AI was most useful when it reduced effort—helping describe an issue or organise records—while leaving its output open to review.
Above all, we learned to design around the outcome. Recording that someone acted is useful; checking whether the action worked is what completes the loop.
What's next for LocalLoop
Our next step is AI-assisted local news mapping: identifying where reported events happened, retaining original sources and reviewing uncertain locations before publication.
We also want to replace demo accounts and local storage with production authentication, a persistent database and secure image storage, alongside stronger moderation and abuse prevention.
Further development will focus on accessibility, clearer progress updates and better duplicate matching. We would also like to test LocalLoop with residents and council teams to understand what evidence is useful and where the reporting process creates unnecessary work.
Our goal is to make LocalLoop a place where people can stay informed, take part and see the results of their contributions.
Shared places. Shared responsibility. A better city, together.
Built With
- applereader
- carto
- gemini
- jszip
- maplibre
- next.js
- onnx
- openstreetmap
- pdf-lib
- react
- rt-detr
- sharp
- tailwindcss
- typescript
- ultraface
- vite
