Thanh Tran Quang
AI Engineer · Edge AI and Computer Vision
One frame, from sensor to device
Capture
A camera sends one raw frame. Real sensors add grain and lose detail in poor light, so the model has to work with input like this.
Detect
The model finds three objects and draws a box around each one. The number on each box is the model's confidence.
Compress
I convert the weights from 32-bit floats to 8-bit integers. The goal is a model about four times smaller that still finds the same objects.
Deploy
The smaller model runs on the device, next to the camera. Frames never leave the device, and each one takes less time to process.
Capture
A camera sends one raw frame. Real sensors add grain and lose detail in poor light, so the model has to work with input like this.
Detect
The model finds three objects and draws a box around each one. The number on each box is the model's confidence.
Compress
I convert the weights from 32-bit floats to 8-bit integers. The goal is a model about four times smaller that still finds the same objects.
Deploy
The smaller model runs on the device, next to the camera. Frames never leave the device, and each one takes less time to process.
What I work on
Vision
- Detection
- Segmentation
- Tracking
- Anomaly detection
Model optimization
- Quantization
- Pruning
- Export and runtime conversion
Deployment
- Edge devices
- Embedded
- Real-time pipelines
Beyond vision
- LLM and RAG apps
- Data pipelines
- MLOps basics
How I work
Measure first
Before I change anything, I measure the model on the device it will run on: speed, memory, and accuracy on real data. Those numbers decide what to work on first.
Ship the smallest model that meets the spec
I start from the requirement, such as a frame rate or a memory limit, and pick the smallest model that meets it. A smaller model costs less to run and is easier to update.
Test on the target hardware
A model that runs well on a workstation can be slow on a small board. I test on the target device early and keep testing there until release.