Inspiration
Every group event we go to ends the same way. Someone says "send me those pics," our group chat fills with half the photos, a few of us AirDrop to each other, and the best shot of us lives forever on someone else's camera roll. Google Drive requires everyone to have enough space in their Drive to upload pictures (shoutout Emily), not everyone in our friend group has an iPhone so we can't use a shared album (shoutout Albert 6 months ago), and no one wants to download a separate app (all of us tbh). And even when we do manage to collect everything, finding the photos we're actually in means scrolling through hundreds of blurry crowd shots.
We wanted something as easy as tossing photos into a bin. Scan a code, dump our pictures, find ourselves, and let the whole thing disappear after the weekend. That became dump, and each event room is a dumpster.
What it does
dump is our mobile web app for collecting photos and videos from everyone at an event, with nothing to install and no account to make.
- Open a dumpster in seconds. We can create one from the site or by texting our iMessage line. We get a link and QR code to share, plus a one-tap Copy Dumpster Link button.
- Drop everything in. Our friends upload photos and videos straight from their phone's browser. If the signal drops, the upload retries on its own.
- Find ourselves. We sign in with a name and a quick selfie. Every person in the dumpster shows up in a face gallery, and tapping a face shows only the photos and videos they're in.
- Search anything. One search bar covers objects, text written in photos, places, and words spoken in videos. We can also tap Ask the dump to describe what we're looking for in plain words.
- Get texted when we show up. We can opt in to an iMessage when new photos of us land or when someone says our name in a video.
- Moments and a map. Shots taken around the same time and place are bundled into moments and pinned on a map. Duplicate uploads are hidden automatically.
- Stickers. We can cut a person or object out of any shot and add it to the dumpster's sticker pack for iMessage.
- The stage. A live view for a TV or projector at the venue that fills up as photos come in.
- The Dumpster Fire. We can react with any emoji, but our custom Dumpster Fire reaction powers a leaderboard of the event's funniest moments, ranked entirely by our group's taps.
- Gone when it's over. Each dumpster burns after 24, 48, or 72 hours, and everything in it is deleted.
Our dumpsters have no hosts or admins. Everyone has the same power, including the ability to delete any photo or video.
How we built it
Frontend. We built dump as a Next.js 15 app with Tailwind CSS, installable as a PWA.
Dumpster links. There are no accounts. Each dumpster gets a long random id that can't be guessed, so the link is the lock.
Upload pipeline. A background worker processes every upload: it strips camera metadata, makes web-friendly versions, flags duplicates, and runs our models before the shot appears for everyone.
AI models.
- Perceptron is our image understanding service. When someone uploads a photo, the background worker sends it a processed copy and gets back a caption, scene tags, visible objects, text in the image, and the locations of faces. We use those results to organize and search photos.
- InsightFace detects and groups faces for the people gallery and selfie matching. Selfies are thrown away after sign-in, and face data never reaches a browser.
- faster-whisper transcribes videos so spoken words are searchable.
- BiRefNet cuts out stickers on our own server.
- Backboard powers Ask the dump and replies to free-form iMessage questions. It answers using the dumpster's photo captions, tags, detected objects, visible text, and video transcripts. Each dumpster gets its own Backboard assistant, which is deleted when the dumpster burns. Face vectors and phone numbers are never sent to it.
Messaging (Photon 2.0). Photon connects dump to iMessage. From our homepage, we enter a phone number, get assigned a line, and Messages opens with a create command ready to send. From there, Photon:
- Starts group dumpsters from a text conversation, either with a fixed command like
create dumpster "Mt. Si Hike" 48hor a casual message like "open one for the hike, two days," which Backboard interprets. - Handles replies in the same conversation, plus sign-in messages using a one-time code that expires in ten minutes and only works from the number it was sent to.
- Sends opt-in alerts: "new photos of you" (batched about every 90 seconds so we're not spammed), a text when someone says our name in a video, and one warning about two hours before the dumpster burns, with the link.
- Uses an idempotency key on every send so a retry never double-texts, and never stores a sender's number on the dumpster or passes it to Backboard.
Dumpster Fire ranking. Users can react to each photo with emojis, including our special “dumpster fire”, creating our dumpster fire leaderboard. Shots are ranked by their Dumpster Fire count, with other reactions only as a tiebreak.
Data and infrastructure.
- MongoDB holds what the app needs to work right now: dumpsters, guests, media records, face groups, stickers, reactions, and expiration times. Every collection has automatic expiration, so data is deleted even if our worker goes down.
- TigerData records events about that activity: uploads, views, reactions, and cleanup runs. It powers our uploads-per-minute stats and the cleanup audit trail.
- Cloudflare R2 stores the photos and videos.
- Vultr hosts our production backend on a single server running a Docker Compose stack: the Next.js website, the background worker that processes uploads, sends alerts, and burns expired dumpsters, our private face matching service, and our speech transcription service. Caddy sits in front to handle HTTPS, accept uploads up to 150MB, and keep the live photo stream flowing without buffering. The face and transcription services have no public port, so only our own app can reach them. We deploy updates with Docker Compose and run health checks to confirm the site and every service come back up.
- Domain. We got dumpster.photo through the MLH GoDaddy link, which redirected us to Porkbun to register it. We then pointed it to Cloudflare to manage DNS.
Privacy and AI auditing (VEYZL). Our privacy rules are checked on every pull request. The build fails if retention isn't 24/48/72 hours, if data could outlive its dumpster, if face data could reach the browser, or if any creator or uploader identity is stored. We also pointed VEYZL's audits at our public vision endpoint. Based on what it flagged, we added clearer refusals, redaction of common personal details like emails and phone numbers, a second check that every answer is safe and supported by the image, and explanations of what information is missing when a question can't be answered.
Challenges we ran into
- Getting VEYZL to reach our model. At first, VEYZL couldn't find or connect to our AI endpoint, so its early audits couldn't test it properly. Once it could, it flagged answers that made unsupported claims, missed instructions, or handled sensitive information poorly.
- A misleading "PII leak." One flagged leak turned out to be a wording issue, not a real leak. The model hadn't revealed anyone's details, but it said it had no customer contacts instead of explicitly refusing the request. VEYZL also wanted a more helpful answer when an identity check couldn't be verified from the image. After we reworked the endpoint, a later audit passed those probes, though it still saw a few intermittent model errors.
- Anonymous but notifiable. We never record who creates a dumpster or uploads a shot, but face alerts still need a selfie and phone number. Keeping that opt-in data scoped to one dumpster and deleted with it took careful design.
- Deletion that actually deletes. Our files, records, face data, and Backboard assistants live in different places, so we built a cleanup worker with database-level expiration as a backstop.
- Faces that drift. The same person at a different angle kept showing up as a new person until we started keeping multiple examples per face.
- Never blocking on AI. We made every model optional, so photos still publish even if one goes down.
- Layering visuals over UI. Our bounding boxes initially covered labels and buttons until we enforced a strict layering order.
- Writing for our users, not for ourselves as engineers. We rewrote our copy so it describes what dump does, not how it works.
Accomplishments that we're proud of
- Going from scanning a QR code to uploading photos takes seconds, with nothing to download.
- Our single search bar finds a face, an object, a word on a banner, or a name said in a video.
- Stickers are made on our own server, so photos aren't sent anywhere to create them.
- Our leaderboard is fully community-driven, so the Dumpster Fire reflects what our group found funny.
- Our vision endpoint went from failing VEYZL's safety probes to passing them before submission.
- We built a visual identity that feels like ours rather than a template.
- We integrated Perceptron, Photon, VEYZL, MongoDB, Vultr, Backboard, and TigerData into one coherent product in a single weekend, live at our own domain.
What we learned
- How to build a background media pipeline that keeps our UI fast while heavy AI work happens behind the scenes.
- That privacy is easier to design from the start than to bolt on. Not collecting host or uploader identity made many later decisions simpler.
- That not leaking data isn't enough. A safe AI answer also needs to refuse clearly and explain what's missing, or it still reads as a failure.
- How to combine hosted and self-hosted models, and when keeping data on our own machines is worth it.
- How much design restraint matters: removing jargon and generic UI patterns made our app feel more human.
What's next for dump
- Recap texts: sending a highlight digest of the Dumpster Fire winners before the dumpster burns.
- Export before expiry: letting each of us save our own photos in one tap before the dumpster disappears.
- Steadier model responses: tracking down the intermittent errors from VEYZL's latest audit.
Built With
- backboard
- birefnet
- cloudflare
- faster-whisper
- insightface
- mongodb
- nextjs
- perceptron
- photon
- porkbun
- tailwind
- tigerdata
- vezyl
- vultr