Campus Lost and Found with Visual Search

The problem with campus lost property

On most campuses lost property is a cupboard in the admin block and, at best, a spreadsheet or a WhatsApp group full of blurry photos. Someone who lost a set of keys or a calculator has no way to search except walking in and asking. Staff have no way to match a description to a box of items, and unclaimed items pile up until someone throws them out.

The same pattern exists in hostels, libraries, sports complexes and schools. The volume is small, but the frustration is real and the fix is cheap.

What already exists and the gap

• Boomerang uses machine learning to match pictures and descriptions of lost items, selling to stadiums, universities, offices and theme parks.
• Crowdfind (now alongside Pixit) lets airports and venues post photos of unclaimed items that travellers can search.
• Chargerback is used by airports, airlines, hotels and parks to manage inventory, claims and shipping.
• Student GitHub projects implement AI matching for campuses, usually without any measurement of how well the matching works.

The commercial tools are built and priced for airports and large venues that handle high volumes and ship items back. A university office needs something much smaller: fast intake, search that works from a description, and a claim check so the item goes to the right person. Most student versions skip the evaluation. You build the small tool and measure it.

What you build

Three roles:

• Staff: log found items, answer claims, hand items over, run the monthly report.
• Student or visitor: searches by description and filters, and submits a claim.
• Admin: sets categories, locations and retention periods.

Main flows: intake, search, claim with verification, handover, archive.

The technically interesting part

Searching photos by text without training anything. A joint image and text embedding model such as CLIP places a photo of a black umbrella and the sentence "black umbrella with a wooden handle" near each other in the same vector space. Store image vectors in PostgreSQL with pgvector, embed the query at search time, and rank by similarity combined with filters. Show your supervisor the ranking, the failure cases (similar looking chargers, glare on phone screens), and how filters fix some of them.

Verification. Anyone can claim an item they saw in a public list. Staff record a private note at intake (the sticker on the laptop, the name on the ID card case, the lock screen image) and the claimant must describe it. Public listings should show a cropped or blurred photo so they do not give the answer away.

How to evaluate it

• Retrieval accuracy: build at least 50 pairs of photos and independent written descriptions from other people. Report top one and top five accuracy, and compare two embedding models.
• With and without filters: show how category and date filters change top five accuracy.
• Intake speed: time staff logging ten items; target under a minute each.
• Pilot: run it for two to four weeks with a real office or hostel and report items logged, searches, claims and returns.

Suggested plan

1. Week 1: talk to the office that holds lost property; collect categories and their current process.
2. Weeks 2 to 4: data model, intake screen, storage for photos.
3. Weeks 5 to 7: embedding service, pgvector search, filters.
4. Weeks 8 to 10: claim workflow, verification, retention and reports.
5. Weeks 11 to 14: evaluation set, model comparison, pilot and write up.

Path to a product

Your first customer is your own university admin office or a hostel. Offer it free for a term, then a small annual licence per campus. Schools, gyms, co-working spaces and event venues have the same problem at a similar scale, so a simple hosted version with per site pricing is realistic. Keep it cheap to run: embeddings are computed once per item.

Scope warning

Do not add in-app messaging, accounts for every student, or a native mobile app. Do not fine-tune a vision model; off the shelf embeddings are enough for this scale. One responsive web interface with staff and public views, done properly, is the project.

Learn what you need on Sefism

• Machine learning roadmap : enough background to explain embeddings and similarity search.
• AI engineering cheat sheet : embeddings, vector search and evaluation in quick reference form.
• Full-stack roadmap : the web app around the model.
• PostgreSQL Documentation : indexes and queries you will combine with pgvector.
• Google Colab : a free GPU to compare embedding models before you deploy.

Get help with this idea

This is a good smaller project if you take the evaluation seriously. Tauseef can help you plan the retrieval test set, choose between embedding models, and keep the scope finishable for one or two people. If you are new here, create a free account first.

• Discuss this idea with a mentor
• Send a query to Tauseef
• Book a 1:1 session