Project information

• Category: MLOps
• Client: Waletech (INSTAR Germany) Shenzhen, China
• Project date: 01 Feb, 2026
• Project URL: INSTAR Cloud

The INSTAR Cloud AI Pipeline initially was build distributed over several Nvidia GPU servers using custom Tensorflow and pyTorch models. To cut cost - while improving the accuracy of the detection - I was task to move everything to TWO servers with updated GPUs. By updating the models and exporting them to CPU friendly format (RTLite and OpenVino) and building multi-threaded application using Redis queues and Python Ray Actors I was able to scaffold a SINGLE CPU server pipeline that could handle the entire payload that was later by me and the Cloud data and frontend team extended to TWO CPU servers for redundancy - still at a fraction of the cost and already battle-proven with 10.000+ detections task a minute queues.