AI Vision Box · Turn existing cameras into AI sentinels · On-device · Scene packs
An RK3588-powered (6-TOPS NPU) multi-channel AI video analytics box: RTSP connection to your existing IP cameras, on-device inference, and 6 ready-made algorithms sold as scene packs — mask / chef hat / flame / smoke / helmet / rodent detection for kitchens, fire safety and construction sites. No camera replacement, no cloud.
- 🎯Crowdfunding goal from ¥20,000
- 💰Box from ¥2,199 · Scene packs from ¥299
What it brings you
Works with your cameras
RTSP connection to existing IP cameras — no camera replacement, no cloud video streaming; auto reconnect and backpressure frame-dropping keep it smooth.
6 ready-made algorithms
Mask 0.90 / chef hat 0.88 / flame 0.89 / smoke 0.85 / helmet 0.91 / rodent 0.86 (baseline mAP, from the LingJing Trainer model store).
6-TOPS edge inference
RK3588 triple-core NPU with ONNX → INT8 quantization under accuracy gates: models exceeding 3-point mAP drop are auto-requantized, never shipped.
Scene packs
Kitchen-compliance pack / fire flame-smoke pack / construction-safety pack / property-sanitation pack — one box one pack, algorithms updated over the air.
Per-SN licensing
Algorithms licensed per device serial with sha256-verified manifests — anti-diversion and remote license transfer supported.
Complete channel system
Activation / 7-day no-questions return / 12-month warranty / reseller pricing with monthly rebates — verified end-to-end on the device platform.
Transparent R&D progress
The percentage is assessed from completed milestones across the whole development plan, and updated as development moves forward.
✅ Completed
- ✅Device management platform MVP running: all 32 end-to-end smoke tests passing (login → inventory → anti-diversion → order → shipping → activation → tickets → swap → analytics)
- ✅6 vision algorithms registered with baseline metrics (mAP 0.85–0.91); YOLO-family decode and 640×640 inference pipeline code complete
- ✅Model conversion pipeline + RK3588 board benchmark code complete; rknn-toolkit2 2.1.0 toolchain locked
- ✅Activation / licensing / OTA API contracts and 14 database tables finalized
🚧 In progress / next
- ⬜Model conversion for real: x86 Linux + real ONNX models producing INT8 device models
- ⬜Board benchmarks on RK3588 SBCs once they arrive, to finalize channel-count specs
- ⬜Box firmware development and production-enclosure thermal soak validation (30-min full-load gate)
- ⬜OTA distribution / e-commerce order sync / WeChat push integration
Key technology
| Hardware | Rockchip RK3588 (6-TOPS triple-core NPU) · 8GB + 64GB · fanless; no built-in cameras — RTSP to your existing IP cameras (PoE edition available) |
|---|---|
| Channels | Designed for 8–16 channels (final count subject to board benchmarks, with 20% headroom before we publish specs) |
| Algorithms | Mask / chef hat / helmet wearing detection, flame / smoke / rodent object detection; 5fps sampling per channel per model |
| Quantization | ONNX → RKNN INT8 with accuracy gates: mAP drop ≤3 points and recall drop ≤5 points required to ship |
| Platform | FastAPI device management platform (activation / licensing / service / resellers / analytics), SQLite switchable to PostgreSQL |
| Scene packs | Kitchen compliance ¥399 · flame & smoke ¥299 · construction safety ¥299 · sanitation ¥349 (reseller pricing available) |
| Box | Multi-channel ¥2,199 · PoE edition ¥2,499, 12-month warranty, 7-day no-questions return before activation |
Back the AI Vision Box crowdfunding
Early access · Founder pricing
Crowdfunding backers get priority access to engineering samples and founder pricing, plus first updates on mass-production schedules.
Co-development · Custom
Enterprises can co-develop on any active project: custom sensors, enclosures, apps and cloud platforms — sharing the R&D outcome.
Channel · Industrial partners
Distributors, system integrators and industrial investors are invited to participate via crowdfunding or strategic partnership.
Do I need to replace my cameras?
What stage is the project at?
Does my video go to the cloud?
We put risks on the table
Crowdfunding means sharing development uncertainty with us — please read these before backing.
- Spec validation: final channel count, thermals and false-alarm rates come from RK3588 board benchmarks; published specs keep 20% headroom, and configuration will be adjusted with public notice if benchmarks fall short.
- Delivery timeline: model conversion, firmware and the production enclosure are still in progress; first deliveries follow milestones, updated publicly.
- Scene adaptation: store lighting and camera angles affect accuracy — installation guides and pre-sales test support provided.
- Push channel: alerts depend on phone notification channels and may delay in rare cases; local siren alerting is on the roadmap.
Got an idea? A machine to make smart?
From a brand-new AI device to retrofitting existing equipment — talk directly with our engineers. Bilingual service in English & Chinese.
📞 Phone / WeChat: 19040667970