Zhang Jia

Project Overview: Enterprise-grade Multi-Agent Runtime platform built on the Orchestrator-Workers architecture, implementing Agent lifecycle management, task orchestration, tool ecosystem integration, context management, and structured artifact generation, enabling multi-Agent collaborative execution in complex business scenarios

Tech Stack: SpringBoot 3 + WebFlux(Reactor) + MCP + Groovy + MyBatis Plus + MySQL + Vue3

Responsibilities:

• Multi-Agent Orchestration Engine: Self-developed state-machine-driven Orchestrator-Workers orchestration engine, implementing task decomposition, sub-Agent scheduling, result review, and terminal state convergence; controlling Agent behavior via tool whitelists, execution round limits, and state guardrails
• Agent Runtime & Event System: Built multi-Agent asynchronous event flow architecture on WebFlux/Reactor, designing a unified Event Protocol to enable real-time interaction for Agent switching, tool invocation, and streaming output via SSE
• MCP Tool Runtime: Implemented an MCP JSON-RPC Client supporting stdio/HTTP dual transport protocols, realizing tool discovery, Schema transformation, and dynamic registration to integrate external MCP Server capabilities into the Agent Tool Runtime
• Context Engineering & Memory System: Built an Agent Context Pipeline implementing role isolation, context trimming, Token Budget control, and Prompt assembly; combining incremental summarization, quality scoring, and long-term memory accumulation for long-session management
• LLM Gateway & Artifact Generation Pipeline: Implemented a model access layer compatible with Anthropic/OpenAI protocols, resolving streaming response protocol differences; improving structured artifact generation stability via Markdown AST and JSON Repair mechanisms
• Skill Runtime: Built a dynamic skill loading framework on GroovyClassLoader, supporting runtime compilation of tool scripts, incremental scanning, and zero-downtime updates for hot extension of Agent capability modules

Project Overview: Enterprise-grade intelligent search and knowledge retrieval platform, building natural language query parsing, search plan generation, multi-source retrieval fusion, and intelligent result generation on LLM Query Understanding, enabling users to query enterprise structured and unstructured data in natural language

Tech Stack: Spring Boot + Elasticsearch + Milvus + FastAPI + gRPC + MySQL + SSE

Responsibilities:

• Concurrent Logic Planning Engine: Developed the Query Plan Tree framework, using LLM to decompose complex Chinese queries into AND/OR/NOT/SEARCH/COUNT logic trees. Executed sub-queries concurrently with thread pools and short-circuit optimization, supporting full-set/difference-set computation and result branch tracing
• Metadata-Driven LLM Routing: Dynamically maintained database field Schema on the Java side while the Python side automatically extracted entities and injected rules, enabling new search dimensions with just one row added to the database — zero code or Prompt changes, solving the pain point of Prompt tightly coupling business knowledge
• Multi-Level Fallback & Dynamic DSL: Designed a three-level fault-tolerant chain of "LLM structured analysis → rule engine → keyword dictionary matching", ensuring stable retrieval even when models malfunction; developed DynamicQueryBuilder for automatic routing and dynamic assembly of ES bool query conditions
• Dual Protocol & Streaming Response Optimization: Adopted HTTP + gRPC dual protocol architecture, serving frontend access and high-performance Java service calls respectively; implemented staged streaming output via SSE, splitting result generation into "deterministic data construction → LLM summarization → citation association" to prevent LLM from directly generating structured factual data and improve perceived responsiveness for long-running tasks
• Microservice Decoupling & Dual-Level Cache: Independently packaged a Chinese relative-time parsing microservice (Time-MCP-server) to uniformly handle fuzzy time expressions; designed Plan/Parse dual-level cache for frequent identical queries, significantly reducing LLM call costs and improving response speed

Project Overview: Enterprise-grade RAG intelligent agent platform, implementing a full-pipeline RAG system from document ingestion to intelligent Q&A. The system fully covers intent recognition, question rewriting, multi-channel retrieval, session memory, and model routing with fault tolerance for production environments. Distributed queue-based rate limiting controls concurrency, while circuit breakers and priority degradation ensure high availability across multi-model scenarios

Tech Stack: SpringBoot + MyBatis Plus + Milvus + Redis + Redisson + Apache Tika + Sa-Token

Responsibilities:

• Multi-Channel Retrieval Architecture: Designed a multi-channel retrieval engine (intent-directed + global vector retrieval + post-processing pipeline), improving retrieval precision while maintaining recall rate
• Question Understanding Optimization: Implemented LLM-based question rewriting and decomposition to automatically complete context and break down complex questions
• Session Memory Management: Designed a session memory compression strategy (sliding window + auto-summary) to reduce token consumption and solve long-context token overflow
• Model Scheduling & Fault Tolerance: Implemented model routing with a three-state circuit breaker, supporting multi-model priority scheduling and automatic failover
• Intent Recognition System: Built a tree-structured multi-level intent recognition architecture that proactively guides users to clarify requirements when confidence is low
• Tool Calling Framework: Implemented an MCP tool integration framework with registry-based auto-discovery, fusing knowledge retrieval with tool invocation
• Distributed Rate Limiting: Implemented a Redis ZSET + Lua sliding window rate limiting scheme supporting global and user-level concurrency control

Project Overview: SaaS short link system providing an efficient, secure, and reliable URL management platform for enterprise and individual users, with in-depth analytics and tracking features for flexible link management and optimization

Tech Stack: Spring Boot + Spring Cloud Alibaba + MySQL + Redis + RocketMQ + ShardingSphere + Sentinel

Responsibilities:

• Message Idempotency Control: Implemented MQ consumption idempotency with Redis, ensuring each message is consumed exactly once within a specified time window
• Cache Penetration Protection: Encapsulated a cache penetration protection scheme using double-checked locking and null-value caching to optimize queries for large volumes of non-existent data
• Bloom Filter Optimization: Used a Redisson Bloom filter for short link existence checks, improving efficiency and effectively reducing invalid query requests
• Traffic Governance & Rate Limiting: Integrated Sentinel for QPS-level rate limiting with graceful degradation to maintain system stability under traffic spikes
• Async Peak-Shaving Architecture: Adopted RocketMQ asynchronous consumption to shave peaks in high-concurrency monitoring data writes
• Cache Consistency Design: Designed the "update DB, delete cache" strategy to maintain consistency between short link cache and database
• Distributed Concurrency Control: Used Redisson distributed read-write locks to support concurrent data modification in high-concurrency short link scenarios
• Sharding & Scaling: Implemented shard routing with ShardingSphere to support targeted queries for redirect and pagination scenarios