Guangzhou Xinhaizhikong Technology Co., Ltd.
Vision that runs on the device, at the site.
The algorithms run on hardware you already have — at the till, beside the line, inside the park. Recognition keeps working when the network drops, and the footage never leaves the site.
Solutions
Retail stores
Recognising goods at the till, checking the kitchen
Barcode-free checkout, shelf condition, morning hygiene checks and behaviour checks — all computed inside the store.
Food service & canteens
Ingredients in, dishes recognised, routines verified
Stock-in and stock-out, dish recognition, bowl-set recognition, and on-site routine checks, judged on the device.
Industrial manufacturing
Inspection beside the line, not after the cloud
Appearance defect detection and workstation SOP compliance. Judgement happens next to the line: no bandwidth, no queue.
Parks & public sites
Hundreds of streams, findable after the fact
Multi-stream ingest, structured storage, and search built for looking back — not for a wall of screens.
Why on-device
Works offline — the network drops, recognition does not
Store broadband stutters, lines are kept off the internet, park networks are isolated. On-device inference does not depend on an outside link: an outage costs you reports, not the transaction in front of you.
Footage stays on site — only results leave
Customer images and process footage are sensitive. Recognition happens locally; only structured results are reported. The raw frames never leave the device on site.
One engine, many scenes — retail, industry and security share the same base
Recognition, matching and retrieval share one on-device engine. Moving to a new scene means swapping a model and a routine — not rebuilding a system.
Products
Algorithms & SDK
On-device recognition SDK
Android and Linux, offline inference, models bound to the local catalogue
Open-source clients
Five business clients
Bakery self-checkout, ingredient management, canteen client, deli counting, face-payment terminal
Hardware
Edge boxes and all-in-ones
Runtime pre-installed; attach a camera and results come out
Platforms
Edge platform + Security PaaS
Devices and algorithms orchestrated from the cloud; the PaaS turns video into structured events — 8,800 streams at a single site, sub-second queries
Customers
Chain bakery — Barcode-free recognition at self-checkout.
Freshly baked goods carry no barcode, store networks are unstable, and the till cannot wait; the model and the catalogue sit on the device.
Precision components — Appearance inspection beside the line.
Reflections, oil film and small defects: fixed imaging, region-by-region detection, per-type thresholds, judged next to the line.
Manufacturing park — Looking back across many streams.
Ingest and time alignment first, then search by event: turning "did it happen" into "which stream, which second".
Contract catering kitchen — Morning checks and routine verification.
Check items become on-device judgements, reported with time and a frame summary; the same day's exception is visible the same day.
Resources
On-device vision deployment for stores
camera placement and mounting height, store network and offline fallback, compute and concurrency, go-live acceptance and operations.
Appearance defect detection on the production line
imaging and lighting, sample and labelling conventions, thresholds and false-positive control, line takt and acceptance.
Video retrieval for site security
multi-stream ingest and the bandwidth budget, storage and retention, structuring and search, private deployment.
Tell us about your site first
Send the scene, a photo of the site, or what you run today. We will tell you plainly whether it is doable, where the risks are, and what it would take. If it is not doable, we will say so.
Contact us