Self-service AI model training platform for industrial and IoT field use. Available now. Goes from data to model in one flow. No algorithm background required.
Use operational feedback data for iterative training and evaluate model changes on separate test data.
Query run status via chat. When creating data sources, collection tasks or training tasks, parameters are shown as cards and only execute after confirmation.
Quick training, CV model training, large model training. Supports single-run standard training and start/stop self-evolving training.
Batch or loop frame extraction from RTSP streams. AI generates candidate boxes for manual review. AI-generated images can fill sample gaps.
Unified management of CV models such as YOLO and DEIMv2, and multimodal large models such as Qwen2.5-VL. One-click conversion to ONNX, TensorRT, OM and BModel.
AIBox handles video input, model inference and result delivery. Its event verification software uses CV to screen candidate events and a large model to check whether they are valid.
Video capacity and performance are confirmed through tests with actual video workloads.
| Model | AIBox LITELightweight edge box | AIBox PROStandard edge box | AIBox MAXHigh-density inference node |
|---|---|---|---|
| Positioning | Lightweight edge box | Standard edge box | High-density inference node |
| Video channels | From 8 channels | 16 channels, core model | 64-channel aggregation |
| Platform strategy | NVIDIA / Qualcomm / domestic Chinese chip platforms | High-performance SoC / edge GPU | x86 + professional GPU / multi-node |
| Typical scenarios | Stores, small campuses, single-point applications | Factories, service halls, campuses | Schools, malls, large campuses |
| Delivery method | Standard product + options | Standard product + options | Standard product + options |
Software support for model deployment and industry applications on chip platforms.
* Product images are illustrative renders; actual product may differ. Video channel counts depend on real video streams, algorithm complexity, end-to-end latency, power consumption and thermal test results. QCS8550 figures are internal test results: W4A16 quantization, 448×448 input, end-to-end latency for a single serial request.
Two product lines sharing one hardware base and voice pipeline: a Cultural Tourism IP Toy for scenic spots and museums, and an AI coding assistant that sits on a developer's desk.
A round-screen badge you can clip to a backpack. Arrive at an attraction and a new stamp appears on the badge. The travel passport app on your phone logs every stamp you collect. The IP character on the screen can also give commentary and answer questions.
Usage scenario screenshot from the concept demo video.
A small round-screen device that sits on a programmer's desk. When AI coding assistants like Codex or Claude Code need confirmation, it alerts you. Approve or reject right on the device.
A press actuator is mounted on the end of the arm to push buttons directly on the fire alarm controller panel. It works by pressing buttons, so the controller itself needs no modification.
You can start with a prototype and validate it using your own video and data before we discuss volume delivery.
