LossnFound

Lost & found matching, but the inventory stays private
Report what you lost — one question at a time
Quick follow-up question from the AI when it’s not sure yet when compared to existing items in the inventory
Where did you lose it? This helps the match
Natural Language date input, parsed intelligently by Gemini
Request submitted, await for response
Admin view: the full list of found items.
Admins add a found item, AI fills in tags/details
Admin: new tickets with the best match score next to each one
Admin opens a ticket and you get the top match with the photo, tags, and why it thinks it’s a match, can approve match

Inspiration

Lost & found at big events is way harder than it should be.

If you’ve ever lost something at a hackathon, you know the drill: you describe it to a volunteer, they dig through a pile, and you hope you get lucky. And if the inventory is public, that creates a different problem — it becomes easier for people to guess details and make false claims.

We wanted a system that feels fast for the person who lost something, but keeps the inventory private by default. That’s the whole idea behind lossnfound.

What it does

lossnfound is a private, AI-native lost & found matcher. It has two roles:

user: submits a lost-item request
admin: logs found items and reviews matches

On the user side, the experience is Typeform-style: one question per screen, smooth transitions, no clutter. You answer:

what you lost
where you lost it
when you lost it
optional photo + contact info

On the admin side, staff add found items (photos, location, notes). Then our matching system tries to connect the dots without ever showing the inventory to users.

We score matches like this:

≥ 0.8: strong match → sent to admin for verification
0.5–0.8: “maybe” → Gemini asks one quick clarification question while the user is still filling the form, we append the answer, then rescore
< 0.5: no match yet → ticket stays open and we notify if something turns up later

One extra thing that matters in real life: people type dates in messy ways (“yesterday night”, “Friday 2pm”, “Jan 12 after lunch”). We use Gemini for intelligent date parsing so “when it was lost” becomes something the system can actually reason about.

How we built it

We kept the stack simple, modern, and scalable:

Next.js for the app (client + admin)
PostgreSQL + Drizzle ORM for the database layer
better-auth for email/password auth and sessions
role-based access control using lowercase roles (user, admin)
Gemini 3 Flash for matching + structured extraction + clarification questions
Typeform UI/UX patterns for the client flow
GSAP + motion/react for that smooth “next question” animation feel
Three.js for extra visual polish (because demos matter)

We built it in a way that supports infinite horizontal scaling: the app is stateless, the database is central, and matching is run as a clean backend workflow.

Challenges we ran into

Getting the Typeform-style flow to feel “effortless” while still collecting enough detail to match accurately
Keeping the privacy story clean: users should feel progress without seeing inventory details
Making the “one clarification question” flow feel natural instead of annoying
Turning messy human time inputs into real timestamps (this is harder than it sounds)
Agentic coding went wrong and we have to refactor a lot of it

Accomplishments that we're proud of

A private-by-default lost & found flow that makes false claims harder
A client experience that feels like Typeform, not a boring form
The “one clarification question → rescore” loop that boosts accuracy without friction
Intelligent date parsing so tickets are actually searchable and comparable
A clean role-based admin system with real auth (not just a fake demo login)
An AI-native architecture where AI isn’t a gimmick — it runs the core workflow

What we learned

UX matters more than people think. If it feels smooth, people trust it more.
The best AI features are the ones that reduce user effort, not add to it.
Dates are chaos. If you want matching to work, you have to normalize time inputs.
“Privacy by design” makes product decisions easier and the story stronger.

What's next for lossnfound

Real Google/GitHub OAuth (we show the buttons right now, but didn’t enable the flows)
Notifications that feel real (email + SMS + admin automation)
Multi-tenant support so different events can run isolated inventories
Stronger pickup verification (QR codes, claim codes, audit logs)
Better analytics for admins (recovery rate, resolution time, hotspots)