Li Changle

I build AI agent platforms and the tooling around them.

Based in Shanghai, China, and open to remote work.

Full-stack / platform engineer at ZhongAn Information Technology

Projects

  1. Brain — AI agent desktop workbench

    Internal product at ZhongAn Information Technology

    The internal agent workbench at ZhongAn Information Technology. I own the LLM observability layer, the shared browser and the design extension; all three are in production.

    • Electron
    • React
    • Node.js
    • TypeScript
    • SQLite
    • Langfuse
    • Go

  2. AI sales practice platform for insurance

    Internal prototype at ZhongAn Information Technology

    A training platform for insurance sales teams: role-play with an AI customer, exams, and AI scoring with human review. Planned and built by me alone; internal prototype.

    • Go
    • huma
    • pgx
    • PostgreSQL
    • React 19
    • TanStack Query
    • Ant Design
    • OpenAPI

  3. Tamora — beauty e-commerce platform

    Live, delivered on my own

    A multi-brand beauty store for the Middle East and Hong Kong. Storefront, API and admin console delivered end to end by me; live, and accepted by the client in July 2026.

    • React 19
    • Express 5
    • TypeScript
    • PostgreSQL
    • Prisma
    • BullMQ
    • Vue 3
    • Naive UI
    • Stripe

  4. ALDEN ROYSTON — luxury watch platform

    Live, three phases delivered on my own

    The site of a Dubai luxury watch brand, in English and Arabic, with the Arabic version laid out right to left. Delivered by me across three phases.

    • React 19
    • Vue 3
    • Spring Boot 3
    • MySQL
    • Redis
    • MinIO
    • i18next

  5. abei — bookkeeping that reads your statements

    Open source, a side project

    Open-source personal bookkeeping. Statements arrive by email, are read into transactions and pre-filled by AI, and a person only confirms. The web app is for people; AI agents get the same capabilities through a CLI.

    • TypeScript
    • Node.js
    • Hono
    • Drizzle
    • PostgreSQL
    • pg-boss
    • zod
    • React

Where I have worked

ZhongAn Information Technology Full-stack / platform engineer 2026.03 —

The technology arm of ZhongAn Insurance, an online-only insurer based in Shanghai.

  • Rebuilt the LLM observability layer of Brain, the internal AI agent desktop workbench, built on Electron, React and Node, with a Go control plane. Traces had been reconstructed from message shapes, which broke sub-agent nesting and lost replies. Replaced it with an event contract and a single collector, with an in-memory upload queue and a SQLite fallback that retries and catches up after restart. In production.
  • Designed a shared Browser Host so a person and an agent work on the same page, tabs and login session: action highlighting, takeover at any moment, per-site authorisation, session isolation, encrypted password storage and Chrome data import.
  • Ported Open Design, an open-source AI design workbench into the product, then rebuilt it as a standalone extension carrying 333 design skills, adding preview-bridge and composer-draft APIs to the extension platform.
  • Built Office and PDF preview and a three-mode Markdown editor (rich text, source, diff). Set up the desktop release and auto-update pipeline and fixed cross-platform issues such as Windows start-up blocking, installer build contamination and non-ASCII paths.
  • Ran the team's LLM gateway end to end through three generations: selection, containerisation, intranet deployment, unified keys, quota and usage. It grew from the 8 people on my team to 40+ users before I handed it over.
  • Built the usage pipeline from session logs through PostgreSQL and dbt to a Lightdash dashboard, and deployed the workbench to Kubernetes with kind and Kustomize.

Anhui Xingrong Technology AI application developer, intern 2026.01 — 2026.03

A software company in Hefei, China.

  • Full-stack work on an AI project-management platform (Spring Boot 3, React 18, PostgreSQL): 37 REST endpoints, 10 tables, 8 screens.
  • Used Spring AI to parse documents and break them into tasks; took part in designing the SSE streaming responses and the human-in-the-loop review step.
  • Took part in evaluating on-premise deployment of GLM-4 and Qwen3 on a 16 × RTX 4090 cluster.

Shangceng Network Backend / full-stack developer, intern 2025.09 — 2026.01

A cryptocurrency trading firm; I worked on the services that feed its trading systems.

  • Price aggregation service in Go against OKX, Jupiter and other exchanges: 50+ tokens polled concurrently every second, 8 million+ calls a day, with local caching, dynamic timeouts and adaptive rate limiting.
  • On-chain signal service that listens for new-token events on BSC and Solana, aggregates them over a five-minute window and pushes the result to the trading system over RabbitMQ.
  • Built the trading review platform on my own from scratch (Next.js 15, TypeScript, React Query, Lightweight Charts for candlestick charts), used by the in-house trading team for daily reviews.

Anhui Sanlian Robotics Java developer, intern (university partnership) 2025.06 — 2025.09

A robotics company in Hefei, China, in partnership with my university.

  • Rebuilt the backend of the company website on Spring Boot, reworked MySQL indexing and async processing, and moved the core flow onto RabbitMQ; peak throughput reached 2,000+ QPS.

Skills and education

Education

Anhui Sanlian University2022.09 — 2026.06

Bachelor's degree, Computer Science and Technology

Certificate

Software Designer, intermediate level

China's national qualification exam for computer and software professionals, 2025

Languages

Chinese (native), English (fluent)

AI and agents
  • MCP
  • Multi-agent integration (Claude, Codex, ACP)
  • Langfuse tracing
  • LLM gateway
  • Dify
  • RAG
  • Prompt engineering
Backend
  • Go (huma, pgx)
  • TypeScript
  • Node.js
  • Java (Spring Boot, MyBatis-Plus, JUC)
Front end and desktop
  • React
  • Next.js
  • Vue 3
  • Tailwind
  • Electron (packaging and auto-update)
Data
  • PostgreSQL
  • MySQL
  • Redis
  • RabbitMQ
  • SQLite
  • MinIO
  • dbt
  • Lightdash
Engineering
  • Docker
  • Kubernetes (kind, Kustomize)
  • CI/CD
  • OpenAPI contract generation
  • Playwright end-to-end tests

About

You can change everything in your life.

My motto

HomeSetting out

In the mist of home, I looked to where the mountains end,and set out on my own,the hills I knew growing farther away.

HefeiGrowing up

From a vocational school in a county town to a university in Hefei.Growing up is the work of my whole life:from school to university, and from university into the world.This city helped me grow, and taught me how things really are.

In my first year of university I was afraid AI would replace backend developers, so I threw myself into embedded systems. Development boards, soldering irons and burnt-out motors soon cost more than my living allowance, and in my second year I went back to Java. That detour left me with one sentence: technology is not valuable; solving problems is valuable.

ShanghaiUnfinished

Where will I end up? Maybe right here, maybe somewhere far away.Either way, the road is under my feet.

I am not a genius, and I would not call myself especially gifted. What I have is flexibility: whatever the job needs, I learn. I wrote Java at a robotics company, Go and Next.js with a crypto trading team, and AI applications at a software company. In March 2026 I moved to Shanghai to join ZhongAn Information Technology, where I build parts of an AI agent desktop workbench.

In my spare time I have built two sites for overseas clients on my own, from the first conversation to launch. I am an introvert and can spend far too long worrying over a single reply. The first time I integrated Stripe, the documentation and forms kept me up at night. I asked AI to research the process, explain it and list the documents I needed, and after a few rounds it stopped feeling impossible.

How I work: I write down the plan and the trade-offs before I build, ask my questions early even if there are a few too many, and leave documentation that someone else can pick up.

Qimen Hefei Shanghai

I believe

  • What people cannot see, they cannot operate. What they cannot operate, they struggle to trust.

    — The Next Step for AI Products Is Not a Bigger Chatbox

  • The model determines what AI can do. The platform determines whether people will trust it enough to act.

    — Why Vertical AI Platforms Need Constraints

  • Redoing work because the direction was wrong costs an order of magnitude more than asking three more questions before starting.

    — A talk to my team

  • Good habits do not come from willpower. They come from structure.

    — A talk to my team

  • The most useful part of taste is knowing what looks bad.

    — A talk to my team

Off hours

Servers
I run three cloud servers. Mail, file storage, a blog, Dify and ComfyUI are deployed with Docker, with Authelia for single sign-on. I wrote a WeChat bot, Hermes: when an upstream project releases a new version, it opens a pull request, builds the image, does a canary release, rolls back on failure and sends the result to WeChat. I am developing SmartIt, which uses AI to troubleshoot several servers; high-risk operations need manual approval.
Notes and new tech
I keep my notes in an Obsidian knowledge base. A server syncs it every night at 22:00 and uses Claude Code to generate a summary of the changes. I study new AI tech such as MCP, skills with a CLI, the pi coding agent and jev. I write Rain Says, short essays on building with AI.
Hardware
I have built small embedded gadgets and a two-legged balancing robot, and plan to build a lightsaber that lights up when swung. I build game scenes in Blender.
Music
Saxophone and bass.
Reading, and IELTS
Recent books: Antifragile, Why We Sleep, Understanding the JVM, Quantitative Trading. Preparing for IELTS.
Games
GTA V, CS:GO, Minecraft.
Downtime
Watching clouds and the sea. I like the seaside.

Writing

Short essays on building with AI, and notes from my side projects.

  1. Project notes

    AI Only Writes Proposals: abei's Contract for AI

    abei lets AI categorize transactions, spot transfers and link refunds, but the only table it can write to is a proposals table. What goes through on its own, what waits for a person, and how any of it can be undone.

  2. Project notes

    Deduplicating Bills Is Not a Yes-or-No Question

    Research for abei, my open-source ledger: how nine products and open-source projects handle duplicate transactions, and the five relationships a matching engine has to tell apart.

  3. Rain Says

    Why Vertical AI Platforms Need Constraints

    A general tool leaves the work of organising AI to the user. A vertical platform decides in advance what AI can see and do.

  4. Rain Says

    The Next Step for AI Products Is Not a Bigger Chatbox

    What people cannot see, they cannot operate. AI needs the same working environment as the people it works with.

  5. Rain Says

    Vibecoding Feels Out of Place, Part 1: Technical Anxiety

    The value was never how elegant the code is. It is whether you can solve the problem in front of you.

  6. Rain Says

    Implementation and Reality

    A solo e-commerce build, from the first client conversation to support after launch. Fixing problems turned out to be the development.

  7. Rain Says

    In the AI Era, Collaboration Is What Really Matters

    Code got faster, but merging did not. AI should also sit between the people who ask for a feature and the people who build it.

  8. Rain Says

    The Compounding Effect of AI Development

    AI has never seen data on how to use AI. Learning to steer it now is a skill that compounds.