AI-native Builder · Agent Systems · Local-first AI · Creative Tools

English · 中文

I build AI-native software around one recurring idea:

Let models reason, let tools execute, and keep the whole system observable and verifiable.

My work spans agent orchestration, local AI runtimes, creative tooling, interactive systems and reproducible research.

I care less about "adding AI to an app" and more about designing software where models and agents are part of the architecture from day one.

Composable Brain × Harness infrastructure for coding agents.

Reasoning frontends and execution harnesses are usually locked inside one product. Agent2LLM separates them and lets you pair them freely — ChatGPT or Claude as the brain, WorkBuddy / Codex / Cursor / Claude Code / DeepSeek Harness as the hands.

Brain   →  reasoning · review · intent
              ↓
          Agent2LLM
              ↓
Harness →  execution · workspace mutation

The core principle is simple: the Brain decides, the Harness acts.

CLI-first, adapter-first, local-first. → nanhudev/agent2llm

Use a web LLM as the 3D brain. Let Blender execute.

Chat2Blend captures Blender Python streamed out of ChatGPT or any other web LLM and pipes it straight into a visible, already-open Blender instance.

Web LLM → Streaming Python → Browser Extension → Local Bridge → Blender Add-on → 3D Model

No manual copy/paste. No API key. No second coding agent rewriting the same bpy code. → nanhudev/chat2blend

A desktop companion that makes coding agents observable.

AgentPet watches what your local coding agents are actually doing and translates real development events into visible desktop behaviour — sessions, file changes, git commits, test results, task state.

It deliberately stays an observer, not a controller. The point is not to drive the agent, but to make its work legible.

→ nanhudev/AgentPet

A local-first voice runtime for AI applications.

Speech recognition, model reasoning and text-to-speech can all run on your own machine, exposed over HTTP / WebSocket.

Usable as a personal voice assistant, a game dialogue backend, an AI character runtime, or the voice layer of another agent. → nanhudev/local-voice-companion

One prompt in. A complete MP4 out.

An agentic video runtime that chains the whole production path:

Research → Script → Storyboard → Render → Voice → Composition → QC → MP4

Available through CLI, Python SDK, REST API, MCP and a local Studio UI. → nanhudev/html-video-workflow

Long-term machine understanding of a person — as infrastructure.

Most "memory" in AI products is just a longer chat history. Cognitive OS separates the layers instead:

• episodic memory
• semantic knowledge
• behavioural patterns
• identity hypotheses
• reflection
• prediction

It asks whether a system can build a stable, evidence-backed model of an individual over time — where every claim carries its supporting evidence. → nanhudev/cognitive-os

AI lesson preparation and classroom simulation for Chinese-language teachers.

One topic produces teaching analysis, lesson plans, presentation slides, a virtual-student classroom simulation and a teaching evaluation.

The design deliberately splits LLM reasoning, hard teaching constraints and visual templates — rather than asking one model to do all three and average 70% on each. Live at bubbleapp.cn/aiteacher.

→ nanhudev/ai-teacher-coach

Turn narrative text into a persistent interactive world.

Extracts characters, locations, objects and relationships from fiction, then maintains an authoritative world state while the player acts in natural language.

Novel → World Extraction → Persistent State → Player Action → AI Adjudication → State Update

→ nanhudev/novel2game-ai

An explainable crypto research terminal.

Built around traceable data, deterministic indicators and explicit model-risk boundaries — it shows what is measured, what is inferred, and where the model is guessing. → nanhudev/tradescope-ai

A reproducible quantitative research sandbox.

Walk-forward simulation with transaction costs, leverage, liquidation and out-of-sample validation. Every result is meant to be re-run, not admired. → nanhudev/quant-research-lab

Large-scale numerical verification, done reproducibly.

An exhaustive OpenCL experiment over negative Collatz trajectories for all starts through 10¹⁰ — longest trajectory found: 1,062 steps, at start −9,177,118,737. Ships with the LaTeX paper and the data needed to check it. → nanhudev/negative-collatz-boundary-1e10

Most of my work comes back to three questions:

1. How should multiple AI systems cooperate? Not one giant agent doing everything — different systems with explicit roles, capabilities and boundaries.

2. How do we keep AI systems observable? Real state, real files, real diffs, real execution, explicit uncertainty. That matters more to me than a convincing-looking demo.

3. What changes when AI becomes part of the runtime? Software designed around models and agents from day one, rather than traditional software with a chatbot bolted on later.

Current focus: agent infrastructure · local-first AI · creative AI · human–AI control boundaries

• Evidence before claims.
• Humans stay in control of consequential decisions.
• A system should say what is real, inferred, or simulated.
• Reproducibility beats an impressive demo.
• AI should cooperate with tools, not pretend to replace them.

余宣均 / Xuanjun Yu — University of Macau. Founder of BubbleLab · bubbleapp.cn

I build AI systems whose claims can be checked.

_{English and Chinese are parallel versions of the same page, not a translation appended below the other.}