CyberFreezeDev
Loading
Small on-site computers that watch, measure and decide; only results reach your dashboard.
The model runs on a small computer beside the camera or sensor, which saves bandwidth, keeps footage on site and works when the internet drops.
Edge AI means the model runs on a small computer at the place where the data is produced, instead of sending everything to a server or a cloud API. A camera on a packing line does not stream video to the cloud: a board beside it detects each carton, counts it and sends one line of data per minute.
That design has practical advantages for factories, warehouses, clinics and shops:
Edge AI and IoT is one of the tracks in CyberFreeze Labs, where we test boards, cameras and sensors before they go into a client design. See the edge AI and IoT lab track.
Each one suits a different size of model and budget. Figures are the vendors' published specifications.
| Board | AI compute | Memory | Power | Good for |
|---|---|---|---|---|
| NVIDIA Jetson Orin Nano Super | 67 TOPS (INT8) | 8 GB LPDDR5, 102 GB/s | 7 to 25 W | Several camera streams, larger detection and segmentation models, small language and vision models |
| Raspberry Pi 5 with AI HAT+ | 13 TOPS (Hailo-8L) or 26 TOPS (Hailo-8) | Pi 5 system memory | Low, Pi 5 power supply | One or two cameras, object detection and pose models, low unit cost |
| Google Coral Edge TPU | 4 TOPS | Uses the host's memory | About 2 TOPS per watt | Adding fast classification or detection to an existing Linux box over USB or M.2 |
| Microcontrollers (ESP32 class) | None needed | Kilobytes | Battery friendly | Reading sensors and sending readings over Wi-Fi or a gateway, where the decision is a threshold, not a model |
Raspberry Pi has also released an AI HAT+ 2 built on the Hailo-10H; we evaluate new boards in the lab before recommending them. Models must be compiled for the specific accelerator, so the board choice is made early.
Units counted and missing caps or dented cartons logged with a photo.
Count units, check a label is present, spot a missing cap or a dented carton, and log each event with a photo for review.
Cold room and fridge sensors alert on WhatsApp and keep a log for auditors.
Temperature and humidity sensors in cold rooms, pharmacy fridges or server rooms, with alerts on WhatsApp when a reading stays out of range, and a daily log you can hand to an auditor.
Current and vibration sensors learn a motor's normal pattern and flag drift early.
Current clamps and vibration sensors on motors and pumps. A small model learns the normal pattern and flags drift before a breakdown.
A camera reads analogue gauges with no data port and logs the values.
A camera pointed at an analogue gauge or a display that has no data port, read by a model every few minutes and written to a spreadsheet or database.
People counts per zone, with no images stored.
People counts per zone for retail and facilities teams, with no images stored, only the numbers.
Lighting, distance, mounting, power and network checked, since placement decides most outcomes.
Photos or a video call of the spot: lighting, distance, mounting, power and network. Most edge projects succeed or fail on camera placement and light, not on the model.
Days of real footage or readings, including glare, night shifts and dirty lenses.
We record real footage or sensor readings from your site for a few days, including the bad cases: glare, night shift, dirty lenses, unusual products.
An existing detector where one fits, otherwise a small model trained on your samples.
An existing detection model where one fits; otherwise we label your samples and train a small model. We report accuracy on footage the model has not seen.
Model quantized for TensorRT, Hailo or the Edge TPU and measured on the device.
Convert and quantize the model for TensorRT, Hailo or the Edge TPU, then measure frames per second and power on the actual device.
One unit in our lab, then one on site for two to four weeks.
One unit running in our lab against recorded footage, then one unit on site for two to four weeks with results checked by your team.
Enclosures, remote updates, a silent-device alert and a results dashboard.
Enclosures, mounting, remote updates, a health check that alerts if a device goes silent, and a dashboard for the results.
Results go to MQTT, a dashboard, a WhatsApp alert or an n8n workflow. We build sensing and vision, not robots or motion control.
An edge device is only useful if someone acts on what it sees. We connect results to the tools you already use: an MQTT broker or a small database on site, a dashboard, a WhatsApp alert to the shift supervisor, or an n8n workflow that raises a ticket and emails a report. When the edge device needs a bigger model than it can hold, it sends the hard cases to a private AI server in your building rather than to a public API.
We do not build robots or motion control. Our edge work is sensing, vision and decisions that a person or an existing control system then acts on.
Often, yes. Most IP cameras offer an RTSP stream that a Jetson or Raspberry Pi can read. Whether the image is good enough depends on resolution, angle and lighting, so we check sample footage first.
Raspberry Pi with an AI HAT+ is cheaper per unit and fine for one or two cameras running a detection model. Jetson Orin costs more but handles several streams, larger models and some small language and vision models. We pick after measuring your model on both when the choice is close.
The device keeps running its model and stores results locally. When the connection returns, it sends the backlog in order. Alerts that cannot wait can go over a 4G modem.
Each device pulls signed updates for its model and software from a server you control, and reports its version and health. A device that stops reporting raises an alert.
Last updated