Projects
Agent systems, engineering projects, and technical explorations.
In DevelopmentClosed SourceAgent SystemsXDU-Zhixing
My role: Primary Harness Developer
An AI project for campus scenarios, with my work focused on the harness that supports agent execution.
Show less
Overview
XDU-Zhixing is an AI platform for the Xidian campus, connecting campus information, search, and community content. Its web app and WeChat mini program share a backend, with an agent harness providing model execution and tool use.
My Contribution
As the primary harness developer, my work focuses on model execution, tool calling, conversation context, and integration with the application backend.
Harness Capabilities
- Search and reading: retrieve campus information, web pages, and academic records; read relevant sources and use community content within the user’s access scope.
- Traceable citations: connect answers to sources actually read, with checks on source availability and access during response generation.
- Streaming and structured presentation: deliver text, tool activity, and status updates, with source lists, step-by-step guides, and service cards.
- Execution and conversation management: handle history budgets, timeouts, cancellation, and concurrency, while persisting messages, sources, and execution traces.
Engineering Focus
The focus is on turning retrieval, models, and tools into useful workflows for campus questions, with attention to source reliability, response experience, and access boundaries. Development is ongoing.
Project Links
This project is closed source and has no public repository or demo link.
In DevelopmentAI Companion ExplorationProject CHIE
An AI companion project built on MaiBot, exploring memory, context management, and behavioral consistency across long-term interactions.
Show less
Overview
Project CHIE is an AI companion project built on the open-source MaiBot. Starting from long-term interactions in group and private chats, it explores how an agent can use conversation history, memory, and context to maintain coherent behavior.
System Components
- Reasoning and tools: a Planner–Replyer workflow separates response decisions from generation and connects tools for memory lookup, replies, and messaging.
- Memory and retrieval: persistent memory organizes conversation records and user information, combining semantic vector search with relational retrieval.
- Interaction and maintenance: platform messaging, a plugin runtime, a FastAPI backend, and a React dashboard support observing and adjusting system behavior.
Ongoing Exploration
Current interests include memory organization, context management, and behavioral consistency over time. Future plans explore separating chie-core from chie-body to decouple reasoning and memory from platform or physical interaction; embodied interaction remains a future direction.
The project preserves attribution to MaiBot. This overview describes the current repository and its research directions; implementation details and upstream attribution are available through the links below.