Soul

Solution

Soul rebuilt its AI delivery layer on Alibaba Cloud Container Service for Kubernetes (ACK) Cloud-Native AI Suite to create a unified internal AI PaaS. The platform covers the full model lifecycle from data management to job release and model evaluation, which gives AI teams one operating surface instead of many ad hoc VM-based pipelines. Unified heterogeneous-resource management automates GPU node handling, separates algorithm code from standard images, and uses auto-scaling inference services so teams can deploy and operate models without repeatedly hand-tuning low-level infrastructure. Soul also uses GPU-sharing mechanisms and Fluid data acceleration to make large AI datasets easier to feed into workloads and to raise resource utilization across the fleet. In practice, the platform has grown from tens of GPUs to nearly a thousand while supporting voice matching, chatbots, OCR, image recognition, and other social-product AI services.

Data flow

Training data and model assets are stored in OSS-backed datasets and accelerated through Fluid, while ACK Cloud-Native AI Suite manages job submission, evaluation, deployment, and inference scaling. AI services such as voice matching, chatbots, OCR, and multimodal workloads then run on shared GPU fleets through the unified platform.