Alan Yu

Agent Infrastructure Engineer

I build AI agents end to end, from the turn loop to long-term memory.

I'm still an explorer. I learn by building: in a field that resets every few months, it's how I keep pace with the frontier, adapt fast, and find what's worth keeping.

Open to work

About

My focus is the harness around an LLM.

How I work:

Own the stack. I've built an agent's core layers by hand — the turn loop, provider protocols (Anthropic Messages, OpenAI-compatible), streaming, prompt caching, context assembly and tool execution — so I know what each costs and how it fails.

Codify the workflow. For each product, I put coding agents under a strict, written workflow — spec, review, test, ship — and every version carries a note on what changed and why, including the ones that were wrong.

Make the product observable. Every agent I ship traces each turn end to end — model calls, tool calls, latency — so a behaviour change is judged on evidence: a test that pins it, or a before/after from the traces.

Keep long-term memory stable. I've explored what keeps an agent's memory reliable over months — layered injection, with each kind of memory entering the context at its own tier, and tuning the scoring parameters of RAG recall so the right memories surface.

Stack

Agent runtime & scheduling
State-machine turn loops with bounded tool iteration · three-state concurrency (safe-parallel / session-serial / global-serial) · conversation state persisted to SQLite and rebuilt on restart · proactive behaviour gated by deterministic rails — quiet hours, idle floor, cooldown, quota · full-turn tracing with a local timeline viewer.
Tool execution & isolation
Zod validation of every tool call's input and output · per-tool timeouts and abort signals, and one ok/err ToolResult · path guards that canonicalize before every check, and an evaluator firewall over the files that gate the agent · SSRF-guarded outbound fetch, pinned to the checked IP · origin-checked transport, loopback by default.
Production engineering
Strongly-typed full-stack TypeScript · one Zod-typed wire contract shared by server and client, so an unmirrored change is a compile error · a supervisor that restarts a crashed backend · every risky feature behind a default-off flag before its default flips · a written rationale per release.

Going deeper on

multi-agent scheduling under contention, stronger isolation boundaries for tool execution (MicroVM and Wasm runtimes), distributed scheduling across agent workers, and durable execution / workflow-replay engines. Direction, not shipped work — the isolation above is in-process.

Languages & runtime

TypeScript (strict) · Python · C++ · Go · Bun · Node · Electron · SQLite · WebSocket · MCP · Anthropic & OpenAI-compatible providers · Git · Docker

Experience

  1. Jun 2026 — Present

    Open-source contributor · airi (~50k★), velin

    airi is a desktop AI companion much like Luna. Users kept asking why the character couldn't look things up online, so I lifted the web layer out of Luna, reshaped it for airi's stage-ui, and shipped it. Also fixed a background-process leak in their computer-use MCP, and landed prompt-as-tool in velin.

    Open source · TypeScript · Vue · MCP

  2. 2025

    Campus Partner, Lead · ByteDance

    Ran hackathons and campus vibe-coding workshops at Chang'an University — teaching students to ship real things with an LLM in the loop. Named Outstanding Campus Partner of the Year.

    Hackathons · Workshops · Developer community

  3. Oct 2024 — Feb 2025

    Technical Intern · E-class, Chang'an University

    Selected 1 of 12 — the interview tasks are still up. Data collection, cleaning and ingestion for AI Counselor, a campus RAG assistant. Working next to the team's engineers is where I first saw how an agent is actually put together — the starting point for everything under Projects.

    Python · Selenium · RAG · Data pipelines

View Full Résumé · PDF

Projects

Luna — a desktop AI companion

She remembers across months, opens conversations herself, and has a face. Three layers of memory over one SQLite file with hybrid recall, and a nightly dream cycle that distills the day into facts, a diary, and skills she keeps for herself. Live2D avatar, phoneme lip-sync, and a cloned voice you can set up without ever opening a terminal — all behind a single Zod-typed WebSocket contract.

TypeScript · Bun · Electron · SQLite · Live2D · Anthropic API

Luna TV — live demo ↗ · Source ↗

agent-kernel

What's left of Luna once you take her personality out: a StateGraph turn loop, a Zod-first tool registry with three-state concurrency, a provider abstraction, and full-turn tracing you can read in a local viewer. Add tools, swap the system prompt, and it's a different agent.

TypeScript · Zod · SQLite

Education

  1. 2023 — 2026

    BEng Artificial Intelligence · Chang'an University

    The first three years of the 3+1, in Xi'an. Everything under Projects was built alongside coursework, on my own time.

    Xi'an

  2. 2026 — Present

    BSc Computer Science · EPITA, Paris

    The Paris year of a Sino-French 3+1 joint programme. I'm in Paris now, looking for internships across France and the EU.

    Paris · Available

Eiffel Tower model: “SM - Eiffel Tower” by Scott Marshall, CC BY 3.0, via Poly Pizza

Looking for a 2026 internship in Paris · English-speaking team · Convention de stage provided