The Engine Room 引擎室

A Chinese SMS classifier that runs on the device

For a messaging company, an on-device classifier for Chinese SMS: a large language model acted as the teacher, labelling examples, and its judgement was distilled into a lightweight student that sorts each message into verification codes, notifications or marketing in well under a millisecond. Three student models were compared before a support-vector machine won on the gold set, with a macro-F1 of 0.91. Fraud detection was deliberately left out of scope.

A thesis: communication is the execution layer for AI agents

In May 2026, in a series of memos and one-pagers, an argument took shape: communication becomes the execution layer for AI agents. An agent that can reach people through the channels they already use — messages, notifications — is an agent that can finish a job. It is a personal thesis rather than company strategy, and it explains why a first-passage mathematician ended up building AI systems inside a messaging company.

AI lead (R&D), Yuexin International

From June 2026, alongside the PhD, AI lead (R&D) at Yuexin International (阅信国际), a cloud-communications company working across messaging, voice and AI. The brief has three parts: the company's AI platform and model gateway; its AI transformation — moving teams from trying AI to working with it every day; and AI support for the international messaging business. The entries below are what got built.

AI lead (R&D)

Feishu MCP connectors: agents inside the workspace

A suite of MCP connectors that lets coding agents work inside the company's Feishu workspace — documents, wiki, sheets, multi-dimensional tables, chats, meeting minutes, approvals and project tracking — acting as the signed-in person rather than as an all-powerful bot. It is a policy layer over the official connector plus custom bridges: per-profile tool allow-lists, secrets held in the system keychain, and a permission tool that automates scope upgrades. Written documentation made it something colleagues could install, not just a demo.

An internal AI gateway with Feishu single sign-on

Within four days an open-source LLM gateway became the company's internal AI platform, the YueXin Intelligent Open Platform. The gateway was chosen by reading the source code of two candidates rather than their brochures. The work covered a deployment runbook, white-labelling and full Chinese localisation of the admin interface, a mobile layout, and a self-service portal where colleagues sign in with Feishu, get a key and see their own usage. Router retries, fallbacks and per-member budgets followed, and a production-grade audit of model and permission boundaries closed the week.

Leading the company's AI transformation

In July 2026 the scattered AI pilots across the company were pulled into one transformation programme: a brainstorm, a working session with the sales department, and then an overall plan with a 90-day operating plan and 30/90/180-day result gates. The first cycle aims at four concrete results — one repeatable internal case, one real external pilot, one working platform loop and one auditable management mechanism — with department-level AI training as the engine.

A morning briefing written by an agent

The sales team asked for more than a daily list of links, so a scheduled agent was built instead. Every morning at 08:30 Beijing time it reads official public procurement notices, extracts the buyer, budget, scope and requirements, connects each notice to the same buyer's earlier notices, and writes a follow-up recommendation into a shared table. Everything with consequences — registering, buying documents, pricing, submitting — stays with people.

An enterprise platform for governed AI access

The company's enterprise AI unified-access and governance platform went from local build to a public site in a week, launching on 31 July 2026. It is a self-built portal over an open-source gateway: organisations, teams and virtual keys, model allow-lists, budgets and usage, a data-loss-prevention hook and an audit hash-chain, with the gateway's own admin interface closed to the internet. Around launch it went through several rounds of adversarial review by parallel AI agents; privacy-canary checks that planted synthetic identifiers and confirmed none survived in storage or logs; and formal production acceptance, with a receipt for every model.

Zijing Career Compass

Zijing Career Compass is an AI career-matching platform for a university's Beijing centre. Students upload a CV, confirm the profile the system builds from it, and see explainable matches against employers on six dimensions, next to a radar of talent policies. The first version was built overnight from meeting notes and went live after a front-end rebuild, followed by a whole-site audit. Later versions added separate workspaces for students, employers and staff, and machine-checked ‘truth gates’ that stop banned wording, personal data and sample placeholders before anything ships. Operations were then handed over to the product owner. The demo runs on synthetic data.

Platform developer

YueXin AI Assistant — web and Windows

In one day, an assistant for company staff was built and launched on the company's model gateway. Colleagues sign in with Feishu and are provisioned automatically, with a personal monthly quota and self-serve API keys. The web app streams its answers, folds away the model's reasoning, renders Markdown and maths, accepts images and has light and dark themes; a Windows desktop client shipped alongside it, with a download page. Before release it went through three parallel security reviews and an independent review by a second model.

From chat window to desktop agent

The desktop assistant is becoming a coding and office agent: an open-source agent engine underneath, a YueXin interface on top. Before the change, three domestic models were shown to complete real coding tasks through the company gateway. The next rounds add quota requests, governance telemetry, a web agent with sandboxed code execution, web search and file reading, a built-in browser, and acting in Feishu on the person's behalf — with a confirmation before every send, share or approval — plus a diff panel and terminal, scheduled tasks, skills and plugins. Telemetry is metadata only, and every employee can see all of it on a ‘My data’ page.