Edge AIComputer VisionKioskPrivacy by Design

Touchless Edge-AI Check-In Kiosk for Any Front Desk

Built for a client in the Netherlands: a walk-up kiosk that recognises people on the device, checks them in and prints a slip on the spot, with no photos stored and nothing sent to the cloud. Works in offices, schools and any staffed entrance.

30 ms

To recognise an enrolled person

Photos stored

On-device

Nothing sent to the cloud

Delivered

Client reports it works very well

Project Gallery

Touchless Edge-AI Check-In Kiosk for Any Front Desk - Image 1

Touchless Edge-AI Check-In Kiosk for Any Front Desk - Image 2

Client countryNetherlands

ClientTechnology company, the Netherlands

DeploymentFront desks: offices, schools and any staffed entrance

TechnologiesEdge AI · Computer vision · Embedded systems · Secure web app

StatusDelivered

ImagesAI-generated illustrations: the real kiosk can't be shown under NDA

The Challenge

At a busy front desk, people queue to be checked in while staff log each arrival by hand.

The client needed a kiosk anyone could simply walk up to, that worked without the internet, kept no photos, and could pass privacy, electrical-safety and accessibility reviews.

The hardware also had to be ready for electrical-safety certification: every power part needed a recognised safety mark, and the printed enclosure had to be flame-retardant.

The Solution

We built the complete kiosk: the enclosure, the AI that recognises enrolled people on the device itself, a slip printer, sound and light cues that guide each person, and a secure web app for front-desk staff.

A person steps up, is recognised, and the kiosk prints their slip and logs the check-in. Nothing leaves the kiosk: images are processed in memory and never saved, and records stay on the device, encrypted.

We engineered for certification from day one: power components chosen only with cULus, CSA or cETL safety marks, a low-voltage DC power circuit bench-tested to the electrical-safety spec, and printed enclosure parts in flame-retardant FR-ABS rated UL94 V-0.

System Architecture

Step 1Walk upA distance sensor, not a camera, wakes the kiosk as a person steps into the lane.
Step 2Recognised on the deviceDetection and matching run on the Jetson in about 30 ms; the frame never leaves memory.
Step 3Slip printed, check-in loggedThe slip prints on the spot and the encrypted log updates for the front desk.

Under the hood

Edge AI & vision

NVIDIA Jetson running TensorRT engines compiled on each unit for its own GPU
About 30 ms per person on the device: face detection ~26 ms + embedding ~4 ms
155° wide-angle camera, fisheye-calibrated so detection sees straight lines
Templates refresh only on high-confidence matches, rate-limited per person

Hardware & certification

Printed enclosure in flame-retardant FR-ABS rated UL94 V-0
Power parts only with cULus, CSA or cETL safety marks
Low-voltage DC power circuit bench-tested to the electrical-safety spec
Slip printer, time-of-flight presence sensor, LED ring, speaker, large button

Reliability & operations

Ready in under 40 seconds from power-on, with no one logging in
Software watchdog restarts a stalled service; hardware watchdog reboots on a hang
Any peripheral can fail without stopping the kiosk
Events queue offline and sync with retry once the network returns

Privacy & software

Zero image storage: frames processed in memory, then dropped
Biometric templates encrypted with AES-256-GCM; key kept apart from the data
Front-desk web app over HTTPS: roles, audit log, bcrypt passwords, idle logout
No cloud services: recognition, records and admin all run on the kiosk

Hard problems we solved

The problemRecognition must feel instant on a small edge deviceHow we solved itDetection and embedding models are compiled to TensorRT on each Jetson and kept warm, so a person is matched in about 30 ms and the slip prints as they arrive.

The problemThe hardware had to be ready for electrical-safety certificationHow we solved itDesigned for certification from day one: power parts only with recognised safety marks (cULus, CSA or cETL) and a DC power circuit bench-tested to spec.

The problemPrinted plastic parts must not feed a fireHow we solved itThe enclosure is printed in flame-retardant FR-ABS rated UL94 V-0, with wall thickness held throughout so the rating stays valid.

The problemPeople of every height, one fixed cameraHow we solved itA 155° lens with fisheye calibration and a software 'digital tilt' crop serves people from wheelchair height to tall adults, with no motorised mount.

The problemFalse triggers from walls, bins and passers-byHow we solved itPresence comes from a time-of-flight beam, not the camera. The kiosk learns the empty lane at power-on and tightens its trigger distance if something sits in range.

The problemFaces change: haircuts, glasses, beardsHow we solved itStored templates blend in new appearance only from high-confidence matches, with a per-person cooldown, so they adapt without drifting towards look-alikes.

The problemA front-desk device must never just stopHow we solved itEvery peripheral is optional at runtime, a watchdog restarts a stalled pipeline, a hardware watchdog reboots on a hang, and events queue until the network is back.

The problemNo photos, everHow we solved itFrames are processed in memory and dropped; only AES-256-GCM-encrypted templates are stored, with the key kept apart from the data.

Key Capabilities

Walk-up check-in: no card, no keypad
Slip printed on the spot
Check-out button for people leaving
Duplicate check-ins flagged
Front-desk web app
Works without internet

Impact

Delivered to a client in the Netherlands, who reports it is working very well
Recognises an enrolled person in about 30 ms, on the device
Stores no photos and sends nothing to the cloud
Delivered with a compliance package for privacy, electrical-safety and accessibility reviews