// Hi, my name is
AI Engineer (LLM / VLM) — China，Shanghai
China Mobile Shanghai Research Institute · Industrial Internet Innovation Dept.
I focus on enterprise-grade LLM and VLM algorithm development — analyzing capability gaps against real business scenarios, then designing the data, training, and evaluation loop that aligns large models to industry needs and ships them to production. My research spans LLM/VLM post-training, retrieval-augmented generation (RAG), multimodal alignment, and personalized federated learning, with papers at NeurIPS, Neural Computation, and Signal Processing.
Targeting industrial natural images, mechanical drawings and time-series charts, analyzed generic-VLM gaps in domain semantics, spatial reasoning and process structure, designed an industrial vision-language alignment scheme and instruction-data system, and ran SFT and iteration to reach >75% image-text Q&A accuracy.
For industrial-procedure Q&A, designed a RAG-based Grounded SFT alignment scheme giving the model evidence-grounded answering and refusal; end-to-end accuracy >84% and faithfulness >95%, shipped to production.
An industrial multimodal understanding system targeting domestic on-device hardware, covering industrial natural images and mechanical drawings; won the Open-Source Impact Award in the Alibaba Tongyi On-Device AI Innovation Challenge.
For industry search, governance and report generation, designed a "CPT + SFT + retrieval-augmented" scheme, built domain training data and an LLM-assisted data-production pipeline, and completed continual pre-training and LoRA tuning; lifted industry Q&A accuracy by 18% and cut data costs by 40%.
Applied split learning to assist training large models (shallow on-device, deep in cloud) and proposed a personalized federated-learning framework introducing a dynamic "federation-collaboration degree"; significant gains across regression, prediction and classification, producing one SCI paper.