Growth Story of a Technical Lead: Building a One‑Stop Large‑Model Training and Inference Platform at Dewu

In Dewu's technology department, the keywords "stability", "efficiency", "experience", "growth" and "innovation" guide the team's culture. Meng, a member of the container technology team, joined the company in October 2022 after working at Tencent, PayPal, Vipshop, Ant Group and Alibaba's DAMO Academy.

He quickly became a benchmark employee by leading the "One‑Stop Large‑Model Training and Inference Platform" project. The platform dramatically reduced the cost of integrating large models and was successfully deployed in community, customer‑service and internal applications, improving business value and user experience.

The interview, part of Dewu's Q2 growth promotion, explores how Meng integrated the concepts of "growth" and "self‑driven" into his daily work.

Project Motivation and ROI

Large models deliver superior performance on complex tasks but require substantial human resources, development cycles and GPU costs. Meng evaluated ROI by focusing on core scenarios (e.g., customer‑service automation), continuously optimizing model performance, and sharing resources across departments. Specific measures included:

Prioritising high‑value use cases to maximize business impact.

Adopting cutting‑edge inference optimizations such as Radix Attention, model quantisation and DeepSeek MTP acceleration.

Building a shared GPU resource pool for training and inference to lower overall cost.

Creating an efficient CI/CD pipeline with one‑click fine‑tuning and deployment, shortening time‑to‑market.

These steps kept the platform cost‑effective while delivering tangible benefits.

Technical Highlights

The platform supports rapid model deployment, Lora fine‑tuning, and integrates with cloud‑native infrastructure. Meng also led the construction of a Milvus vector‑database platform, learning the technology from scratch, contributing to community discussions, and solving stability issues through performance testing and optimisation.

Personal Growth and Self‑Driven Learning

Throughout the project, Meng demonstrated self‑driven growth by:

Learning Milvus theory and practice through documentation and community interaction.

Participating in open‑source discussions, receiving feedback, and applying it to improve system stability.

Sharing experiences at industry conferences, turning challenges into learning opportunities.

He advises newcomers to maintain a growth mindset, seek mentorship, and treat every difficult task as a chance to expand their skill set.

Future Outlook

Looking ahead, Meng plans to focus on further optimisation of large‑model deployment performance, explore Retrieval‑Augmented Generation (RAG) and agent‑based applications, and investigate tighter integration of large models with cloud‑native environments to improve resource scheduling and efficiency.

The interview concludes with a reflection on the importance of meticulous code, iterative design and continuous self‑improvement as the foundation for lasting personal and organisational growth.