University student turning agent ideas into inspectable, stateful systems.
Working on coding agents, role-playing runtimes, and source-grounded knowledge tools.

A local Java coding agent designed for long-running codebase work. Append-only sessions preserve branching history, while contract-driven modules keep models, tools, permissions, subagents, memory, and the terminal UI replaceable.

Java 21 Spring Boot JLine JSONL Maven

Architecture | CI

A modular role-playing runtime built on Pi that keeps storytelling and state transitions separate. World packs compose GM decisions, Writer output, deterministic SQLite state, character agents, snapshots, and optional long-term memory.

TypeScript Node.js SQLite OpenViking

Design notes | CI

• liiy-llm-wiki - A local-first knowledge agent that keeps model output grounded in traceable source material. TypeScript Node.js
• gemini-for-claude-code - A Claude Code plugin for scheduled Gemini jobs and bounded tool-driven code generation. JavaScript Shell
• liiy-filePlatform - A database-free file relay with expiring uploads and a single-binary deployment. Go

My recurring design question: what must remain inspectable when the model, provider, process, or interface changes?

• Recoverability. Append-only history should make work safe to branch and resume.
• Explicit boundaries. Tools, permissions, and state transitions should stay visible and auditable.
• Replaceable edges. Models, interfaces, and storage should sit behind stable contracts.
• Grounded memory. Durable knowledge should lead back to sources, not preserve generated summaries alone.

• Primary languages: Java TypeScript
• Runtime and data: Spring Boot Node.js SQLite JSONL JLine
• Also using: Go Python Shell Maven npm GitHub Actions

Email  /  Repositories