• Experience.

    Work Projects

    • Ant Group · Ant Insurance FinTech 2024.06 - Present

      Served as project owner for MCP and Insurance Quick Check. Since 2026.09, project owner for the Unified Agent Harness Platform (“Da Bao”), coordinating three engineers to build evaluation, experimentation, and release capabilities.

      • Unified Agent Harness Platform (Da Bao) 2026.09 - Present

        Project owner coordinating three engineers to extend an existing early prototype into a unified Agent evaluation, experimentation, and release platform. Now integrated with Ant Xiaobao and Insurance Quick Check, the first end-to-end loop covers monitoring → Bad Case alerting → root-cause analysis → responsibility-domain routing → sandbox validation → release → monitoring; current work continues on Harness architecture, evaluation assets, and release gates.

      • Insurance Quick Check · Architecture / Context / Page AIGC 2025.06 - 2026.08 说明

        Insurance Quick Check architectureProject owner for the Ant Insurance homepage vertical-search entry. Coordinated two application, one algorithm, two search/recommendation, and three frontend engineers across teams; led the architecture and production pipeline across context assets, page AIGC, and Chatbot. A Runtime-Executor-SessionPool Harness supports multiple production workflows and is partly reused by risk-control and post-sales teams. Serves 8M Agent MAU with peak online traffic around 2,000 RPM.

        Produced structured context from policy documents, OCR materials, and product facts for page AIGC and Chatbot use. Across about 20,000 products and 100,000+ tasks, long-running scheduling, automatic retries, and quality validation achieved 100% final delivery success. Over a rolling 200+ sample set, the evaluator measures completeness, faithfulness, logical expression, and compliance at about 98% accuracy. Meta-evaluation uses RCR (Repeat-Consistency Rate) to measure stability, currently about 90%; established ESF (Expert-Score Reproducibility) from scratch, aligning expert judgments by Checkpoint and reaching about 70%.

      • Chatbot · Product-Pool RAG / Embedding / A2UI 2025.06 - 2026.08 说明

        The Chatbot separates product selection from product presentation. Algorithm / search teams own Agentic Search selection on top of my Harness; I owned the product-pool RAG / Embedding pipeline, including query embedding, Elasticsearch retrieval, concurrent fallback, and source navigation over product materials.

        Led the A2UI backend protocol and event stream, validating and finalizing the contract through a full-stack demo. The event stream drives the frontend team’s implementation of interactive cards and shelves; I owned the protocol and backend path, while frontend engineers owned production rendering, enabling more complex and flexible generative UI than fixed interfaces.

      • MCP / Tool Calling Toolchain 2024.12 - 2025.06 说明

        Led MCPBridge and the Ant Insurance domain MCP Server, exposing 100+ existing HTTP / HSF services—business data, rules, retrieval, and analytics—as reusable Agent tools.

        Designed unified Tool Schemas, discovery, argument construction, error feedback, and context enrichment to support the OneAgent + MCPs domain-Agent pattern.

    • Taotian Group · Tao Factory 2021.06 - 2024.06

      Worked on business-finance integration and pre-sales AI, spanning complex data platforms, RAG knowledge bases, intelligent Q&A, recommendation, and business automation.

      • RAG Knowledge Base and Intelligent Q&A 2023.10 - 2024.06 说明

        Knowledge base architectureBuilt a multi-source RAG pipeline integrating merchant, operator, algorithmic recommendation, and missed-query knowledge for full-category pre-sales and after-sales Q&A.

        Designed multi-path retrieval, LLM reranking, answer generation, and missed-query feedback as a continuously improving data-product loop.

      • Money-Bill-Invoice Integrated Data Platform 2022.10 - 2023.10 说明

        Led automation and integration for settlement, billing, and supplier invoicing across 10+ billing scenarios and tens of billions of records using MaxCompute, PostgreSQL, and BigTable.

        Built a domain DSL for billing and accounting rules, reducing siloed development cycles to one week; published as a best practice by Alibaba Technology.

    Open Source

    • Skill.
      • Agent Harness / Runtime: Runtime, Executor, Session Pool, event streams, tool scheduling, long-running task governance, session state, and incident analysis

      • Agent Applications / AIGC: structured context production, offline / online hybrid generation, A2UI protocols, and financial-domain delivery

      • Evaluation / RAG / MCP: generation and meta-evaluation loops, vector retrieval and source navigation, Tool Schemas, and domain MCP Servers

      • Engineering / Data: Python, Java, React, PostgreSQL, Elasticsearch, MaxCompute, Flink, MySQL, and BigTable