AI that runs on the device, not in the cloud.
People counting, product recognition, queue detection or a local assistant — when the analysis has to happen on site, our RK3588 and RK3576 hardware runs it on a 6 TOPS NPU, with public RKNN and RKLLM guides.
Camera in, decision out — on the device.
Video from a camera or an HDMI source goes straight to the NPU, and the model runs locally. Video does not have to leave the site, and the analysis keeps running when the network does not.
- Camera or HDMI inLive video into the device
- NPU6 TOPS on RK3588 and RK3576
- ModelRKNN vision models, RKLLM language models
- ResultCounts, alerts or on-screen content
The checklist we build to.
- An NPU sized for real-time vision
- Small language models running locally
- Camera input and HDMI capture
- Public guides for converting models
- Linux builds as a development project
Where this hardware works.
Common applications by industry, and the setup we would typically build for each. Every project starts from your site, so sizes and parts change with it.
People counting and dwell time
Cameras count visitors, measure dwell time and show how areas are used, without sending video off site.
- Typical setup
- An AI Box next to existing cameras, sending only counts to your dashboard.
- What matters here
- Anonymous counts, processed on the device.
Queue alerts
When a queue grows too long, staff get an alert to open another lane or counter.
- Typical setup
- An RK3588 board or AI Box with an overhead camera.
- What matters here
- Fast enough to react in the moment.
Shelf and stock monitoring
Cameras spot empty shelves and misplaced products and tell staff what to refill.
- Typical setup
- An RK3576 or RK3588 board with shelf cameras; only events leave the device.
- What matters here
- Events instead of video streams.
Content that reacts to the audience
A display notices someone approaching and switches to relevant content, logging only anonymous counts.
- Typical setup
- A display with an RK3588 board and a camera.
- What matters here
- Anonymous counts only — no identification.
Safety zones and protective gear
The camera checks for helmets and vests and warns when someone enters a danger zone.
- Typical setup
- An AI Box or board with the site camera, alarm output over GPIO.
- What matters here
- Local alerts, even when the network drops.
An assistant that works offline
Visitors ask questions by voice or text and a small language model answers on the kiosk itself.
- Typical setup
- An RK3588 running a small model on the NPU with RKLLM.
- What matters here
- Answers without a cloud connection.
One team, from the board to the software.
A typical rollout stitches together a board vendor, a panel supplier, a firmware contractor and a software team. We design the board, build the display and write the firmware ourselves, then integrate your software — so when something has to change, one team changes it.
-
01 BoardDesigned in-house
KTM-S88CD (RK3588) and KTM-S76 (RK3576) with a 6 TOPS NPU; the RK3588 AI Box adds 8K decoding and an HDMI input.
-
02 Display and enclosureBuilt in our factory
The same boards go inside our displays and kiosks, so the AI and the screen ship as one device.
-
03 FirmwareWritten by our team
# RKNN — a vision model on the NPU from rknnlite.api import RKNNLite rknn = RKNNLite() rknn.load_rknn('yolo_rk3588.rknn') rknn.init_runtime( core_mask=RKNNLite.NPU_CORE_AUTO) out = rknn.inference(inputs=[image])Android firmware as standard; Linux builds as a development project under agreement.
-
04 Software and testingIntegrated and burned in
Public RKNN and RKLLM guides in our developer wiki take a model from training to the NPU.
Built to your spec.
Standard products ship as they are. For a project, any of these can change — and because the board, display and firmware are ours, the change is made at the source.
- Camera and sensor integration
- RK3576 or RK3588, sized to your model
- Help converting your model for the NPU
- AI built into a display or kiosk
- A Linux build as a development project
- Enclosure and cooling for the installation
Hardware for this job.
Convert models with RKNN or RKLLM — the steps are in our developer wiki. On-device AI with RKNN and RKLLM
Edge AI & Vision
Can RK3568 boards run AI?
Light models only (about 1 TOPS); choose RK3588 or RK3576 for real-time vision or local language models.
Is Linux available?
The standard firmware is Android; Linux builds are provided as a development project under agreement.
Do we need an internet connection for the AI?
No — inference runs on the NPU. A connection is only needed to send results or update the model.
Which models run on the NPU?
Vision models converted with RKNN — detection, classification and counting — and small language models with RKLLM.
Does on-device AI help with privacy rules?
It helps: video can be processed on the device and only anonymous counts sent on. Compliance with GDPR or other privacy laws still depends on how the system is deployed and used, so check the purpose, notices and retention for your site.