Selected engineering work
Multi-camera retail shelf-vision system
Verified active years: 2024, 2025, 2026
Multi-camera product recognition on YOLOv7/v9 with TensorRT — sub-200 ms responses at 15–20 FPS, hot-reload model serving, end-to-end CI/CD.
The problem
Shelf-vision systems need to recognise products and identify stock gaps across multiple camera views. The engineering challenge was to combine calibration, matching and inference in a responsive pipeline that could also support model updates.
My contribution
AI / ML systems engineer
Engineered multi-camera product-recognition pipelines with YOLOv7/v9, TensorRT and LightGlue. My work included camera calibration, sensor-driven session flows, GPU-optimised execution and CI/CD for model deployment.
Team / platform context
This case study covers my computer-vision and deployment contributions within a retail shelf-vision system. The reported latency and inference throughput come from the existing project record; the synthetic demo is a separate illustration of the concepts.
Try it on this page
Interactive demo
This illustration and the linked demo use synthetic shelf states to explain general engineering concepts. They contain no employer or client software, data or imagery. The preview is a static illustration, not live inference; its shelf states are not production results.
Explore a synthetic shelf scenario
Computer vision · illustrated
Synthetic sceneFrom shelf to signal
Two views. One shelf. A stock state people can act on.
- 01
Capture
Calibrated cameras and sensor-driven sessions.
- 02
Detect
YOLOv7/v9 with TensorRT for product recognition.
- 03
Match views
LightGlue and calibration connect overlapping observations.
- 04
Surface the signal
Stock-state information supports replenishment decisions.
Connect camera observations to stock information
I worked on real-time product recognition and out-of-stock analytics across retail shelf environments. The system combined YOLOv7/v9 detection, TensorRT acceleration and LightGlue feature matching.
Connect multiple views
My work included high-precision camera calibration and multi-camera processing. Matching and calibration provide the relationship between views that a detection model alone does not establish.
Work on response time and delivery
I refined sensor-driven session flows and GPU-optimised execution, delivering the documented sub-200 ms response times and 15–20 FPS inference. I also worked on end-to-end CI/CD for model deployment.
Separate the illustration from the implementation
The shelf illustration explains stock states, and the interactive demo lets a reader explore synthetic scenarios. Neither contains production models, camera imagery or client data.
Engineering decisions
Optimise inference within the full pipeline
The implementation combined YOLOv7/v9 with TensorRT acceleration and GPU-optimised execution. Sensor-driven session flows were also refined as part of delivering the documented response time and throughput.
Treat camera matching as an engineering concern
High-precision calibration and LightGlue feature matching supported the multi-camera workflow. This addresses the relationship between observations from different views alongside the detection model itself.
Include model delivery in the engineering scope
End-to-end CI/CD supported model deployment. This makes the delivery process part of the system being engineered, alongside inference and analytics.
Outcomes & evidence
- Delivered sub-200 ms response times and 15–20 FPS inference in the documented shelf-vision pipeline.
- Contributed multi-camera calibration and matching, sensor-driven session flows and GPU-optimised execution.
- Implemented end-to-end CI/CD for model deployment.
Latency and throughput figures come from the existing project record and public resume. Hardware, camera workload and latency measurement boundaries are not documented publicly here, so these figures are not a reproducible end-to-end benchmark. The demo uses synthetic shelf states and performs no model inference.
Engineering takeaway
Inference speed is one part of a usable vision system. Calibration, matching, session handling and deployment all contribute to turning camera output into information that can be acted on.