FR101 Industrial Edge AI Webinar Demo

This demo is intended to illustrate the possibilties of SLMs on QC6490 based devices running in combination with accelerated Edge Impulse models. The demo is tested on onlogic FR101 but can be run on any 6490 based device locally or on Device Cloud

After the FOMO-AD model triggers on a threshold the reult can passed to be analysed by the SLM for the operator:

This cascading can further be enhanced and extened to VLMs

Telegram interaction

Demo stack for the OnLogic FR101 / Qualcomm QCS6490 webinar.

Flow:

Razer Kiyo camera
  -> Edge Impulse crack / FOMO-AD .eim
  -> repeated anomaly threshold
  -> GPIO / DIO dry-run or output
  -> local Qwen 0.5B / 500M operator note

Recommended live path

Use CPU .eim + local Qwen for the stable demo:

cd /home/onlogic/webinar
./scripts/start_qwen.sh

In another terminal:

cd /home/onlogic/webinar
./scripts/start_fr101_demo_cpu_qwen.sh

Open:

http://192.168.1.58:8080

Model files to copy manually

Large .eim and .gguf files are not included.

models/fomo_ad_cpu.eim
models/fomo_ad_qnn.eim
llm/qwen/qwen2.5-0.5b-instruct-q4_k_m.gguf

Camera

The Razer Kiyo has appeared as /dev/video2 on the FR101. Check with:

v4l2-ctl --list-devices

If the node changes, run with:

CAMERA_DEVICE=/dev/videoX ./scripts/start_fr101_demo_cpu_qwen.sh

Threshold

The Visual GMM project metadata showed min_anomaly_score: 6.0. The app defaults to 6.0, but you can override:

THRESHOLD=0.8 ./scripts/start_fr101_demo_cpu_qwen.sh

QNN path

After installing QAIRT:

source ./qairt-env.sh
ldd models/fomo_ad_qnn.eim | grep "not found" || true
./scripts/start_fr101_demo_qnn_qwen.sh

If the QNN model initializes but classify() returns error -3, use CPU as the live fallback and treat QNN as runtime-validation evidence.