Designed and implemented 5 microservices (PPT parsing, document parsing, speech synthesis, video matting, video compositing) + HeyGem digital human driving module, decoupled via Redis message queue
Implemented HeyGem Broker-Worker architecture, supporting 24 Pod parallel GPU inference with 6 concurrent digital human generations per task
Completed K8s (TKE) deployment with qGPU memory slicing for multi-Pod scheduling on single GPU
Video compositing pipeline: MiniMax TTS → digital human lip-sync driving → green screen matting overlay → multi-page concatenation + Jianying draft generation
Implemented PPT animation splitting: supports [ANIM] markers to trigger animation steps, preserving native animation effects in PPT-to-video conversion
Developed mov transparent channel video processing: alpha hard-cut to pure green → HeyGem lip-sync driving → precise matting with no green edge residue
Integrated power_point_api: user's local Agent remotely calls PPT-to-PDF/video, with automatic fallback to LibreOffice on failure