AI models that run on your device, in real time
Model conversion, quantisation and optimisation for Rockchip NPUs, NXP, Jetson and ESP32-S3, with full camera-to-result pipelines and measured accuracy.
- RKNN
- TensorRT
- TFLite
- ESP-DL
- ONNX
- INT8
What you get
A working on-device pipeline
Capture, pre-processing, inference and output, profiled end to end.
Accuracy you can trust
Quantisation with your real data and side-by-side checks against the original model.
The right hardware choice
Benchmarks on candidate chips before you commit to a design.
Problems I solve
If one of these sounds familiar, we should talk.
- The model is accurate on a PC but wrong on the device
- Inference is too slow for real time
- The hardware is too expensive for the business case
- You need AI without sending data to the cloud
- You are not sure which NPU chip to choose
- A demo deadline is coming and the pipeline is not ready
Questions
Which models can you deploy?
Detection, classification, segmentation, tracking and anomaly detection models from common frameworks.
Can AI run on a microcontroller?
Small, quantised models can, for example on an ESP32-S3. I benchmark before recommending it.
Do you train models too?
I retrain and adapt models for the target hardware and your data.
Have a board that needs to work?
Send your SoC or MCU and the goal. Reply within one working day.
Send a project brief