Returning sports equipment,
verified by AI in seconds.
Transparency: Equip Sport is where our founders work. This system was built by us end to end, and the numbers below come from its production dashboards.
Every photo, by hand, every day.
Equip runs 500+ smart sports-equipment stations across Europe and North America. Players unlock a station with their phone, take what they need for a game, then return it the same way.
Every return ends with a photo of the locker. Until last year, someone on the operations team opened those photos at the end of every working day, station by station, scrolling back from the most recent return looking for anything missing.
When something was missing (a basketball, a paddle, a volleyball), they walked back through the reservation log to figure out which player had taken it last and never brought it back.
It worked. It didn't scale. As station count and return volume grew, the backlog grew with it; turnaround on a missing item slipped from same-day to several days, and ops time on the queue kept climbing.
A two-stage pipeline that checks itself.
Every return photo runs through a small AI pipeline before it ever reaches a human. The overwhelming majority are decided automatically; only the genuinely ambiguous ones land in a review queue.
The pipeline runs in four steps. First, a return photo is captured by the station. Second, an object detection model finds the station and each individual locker, drawing bounding boxes. Third, a vision LLM compares each cropped locker against the most recent maintenance reference frame to decide whether equipment is present or missing, with a confidence score. Fourth, all-present high-confidence results auto-approve the return; ambiguous cases route to a short human review queue.
Detect. An object detection model finds the station and each individual locker in the return photo, drawing a tight bounding box around every compartment.
Verify. Each cropped locker is sent to a vision LLM along with the most recent maintenance reference, a known-good "all equipment present" frame from the last time the station was serviced. The model decides present or missing for every slot and returns a confidence score.
Decide. When every locker comes back present with high confidence, the return is auto-approved and the station is good to go. Anything ambiguous (a low-confidence call, a locker the detector wasn't sure about, an unfamiliar reference) drops into a human review queue that's now a fraction of its old size.
The maintenance reference refreshes whenever a station is serviced, so the model is always comparing today's return to today's known-good baseline rather than a stale frame from months ago.
A long-tail human queue, automated.
The pipeline now handles the overwhelming majority of returns without anyone looking at them. The ops team's weekly review went from a recurring multi-hour scroll to a short triage of genuinely ambiguous cases.
When something is missing, it surfaces inside the same day instead of waiting for a human pass at end-of-shift, so the reservation-log lookup happens while the trail is still warm. That means the person who took it gets contacted before the next time they try to use the system.
The system is live across the full Equip footprint and processes roughly 40,000 photos a month.
Want one of these?
Tell us where you're starting from and we'll point you at the right way in: a deep-dive build, a focused feature, or a quick production-readiness pass.