Owned a cross-industry AI Agent customer-service SaaS platform from architecture evolution through production operations, covering 2k+ tenants, 230k+ users, and 60k peak requests/day across Runtime, tools and Skills, RAG/Memory, quality evaluation, cost governance, and team delivery.

• Led model-driven Agent Runtime 2.0 with WorkflowAgent across Agentic RAG, multi-turn tool calls, termination decisions, and parallel multi-agent orchestration; held end-to-end latency around 7s under 7k+ average conversation-token load, with no-tool scenarios around 1-2k tokens.
• Designed unified context for multi-agent execution, using stable prompt prefixes, deferred dynamic context, role-isolated nodes, and phase-level usage auditing to improve KV cache reuse and pinpoint latency and token drivers.
• Built dynamic Skill routing and on-demand tool mounting, retrieving SOPs and dedicated tools by intent while resolving ambiguous multi-Skill matches through clarification and rerouting to balance context length, tool misuse, and task quality.
• Established end-to-end observability and evaluation across requests, execution phases, tool calls, and terminal callbacks, using correlation IDs, traces, token usage, replay sets, and release gates to isolate quality, latency, and cost issues.
• Built an unknown-question clustering and topic-discovery platform that turned about 10k+ to 60k+ weekly questions into content gaps; about 50% of clusters fed knowledge base/RAG/Skill iteration, high-frequency topic handling reached about 70%, and auto-resolution improved by about 30%.
• Led a 3-person Agent squad through Runtime service decomposition and platform delivery, standardizing gateway authentication, service contracts, container builds, and release conventions; expanded the AI development harness across 40 repositories with Spec/EPIC workflows, acceptance gates, self-check scripts, and PR templates.

Owned web engineering and microservice architecture for AI image generation, model training, and platform products, contributing to 0-to-1 generative AI SaaS delivery.

• Led frontend/backend architecture, domain decomposition, containerized deployment, and DevOps workflows for AIGC products such as miaohua, supporting proprietary image models, LoRA training, and open-source model integration.
• Built model parameter formatting, prompt optimization, task queues, model-service layering, logging, and monitoring around Stable Diffusion / PyTorch inference workflows.
• Designed service boundaries for identity, model inference, model-file storage, and task delivery while supporting cross-platform Web / H5 / mini-program product experience.

Worked on high-traffic Taobao Live and Taobao content scenarios, focusing on micro-frontend orchestration, component communication contracts, SSR/edge rendering, and performance optimization.

• Led the Taobao Live web interaction micro-frontend solution, evolving from isolated Weex components to H5 component orchestration with unified layout, routing, event bus, and communication contracts.

Owned ad delivery, order attribution, cross-platform H5, mini-program, and merchant-console engineering for Alibaba Local Services commercial promotion, connecting consumer traffic with merchant operating data.

• Designed a shell / iframe based micro-frontend architecture to standardize sub-application communication, routing, static resource loading, and deployment.
• Built Vue / React scaffolds, cross-platform JsBridge, ad logging SDK, and reusable components, contributing to key flows from ordering and ad settlement to attribution and merchant analysis.

Owned financial reporting, Java web services, and frontend migration from AngularJS to Vue with a shared Web UI component library.

• Built financial-reporting aggregation and business management systems with Spring Boot, MyBatis, AngularJS, and Vue.