I'm a builder: I get curious, go deep, and build it with AI.
I build the AI that runs a real business, and I go deep into whatever a problem needs. I've worked across operations, sales, and venture.
Most of what's here I built and run solo, on personal subscriptions, with no budget.
I run my client operation on these. Each one does a job I'd otherwise have to hire for, which is how I keep a full client book moving on my own.
Self-updating CRM. My client conversations live on WeChat, which gives you no API, so a background daemon and a Swift accessibility layer reach the data directly and Claude turns each thread into a CRM update, running 24/7 as a self-healing service I review before it commits. The board I built on top brings my WeChat and email pipelines into one view, surfaces what needs action this week, and auto-flags leads going cold so nothing slips.
Email outreach engine. A full LLM email pipeline that researches each company, picks the angle, drafts the email and sends it, then sequences the follow-ups and sends them on a cadence. It tracks opens and replies, and a dashboard turns that into a read on what's landing, so I'm always tightening the copy and the flow. The reliability sits in the engineering around the model: a SQLite state machine, dry-run and validation gates, and a hard cap on daily spend.
RedNote (Xiaohongshu) lead radar. An LLM lead-finder for a platform that bans bots: it scores posts and comments for buyer intent and surfaces the strongest to me. It stays strictly read-only behind a kill switch and an audit log, so it can never post or risk the account. And it closes the loop: the same intel feeds a writing agent that turns what's trending into posts in my voice, covers and all.
Mei, a compliance assistant I can put in front of a client. An LLM assistant for U.S. employment-law questions, designed around trust: it answers only from a sourced knowledge base and current .gov sources, cites every claim, refuses when it can't ground an answer, and ships only after passing a suite of real-question evals.
The judgment underneath. The calls I'm most deliberate about are where AI can act on its own vs. where I stay in the loop, what guardrail each one needs, and what data ever leaves my machine. It all runs on one source of truth, so a fresh chat picks up a client's full context on the spot.
Employee Rights Tool (live · code). A free, no-login web tool that helps a U.S. worker tell whether their situation is illegal, what to save, and how to find a lawyer. I built it after one of my posts on workers' rights took off and the comments made the need impossible to ignore, so I turned the compliance knowledge I sell to employers into something that helps people on the other side of the table. Free to use, and it grows the audience I write for.
Land of Opportunity (play · code). A dark-comedy game about how fast an American small business can die by paperwork. A half-built side project I made for fun; every rule in it is real.
Anthropic API and Claude Code, Python, Playwright, MCP, SQLite, Next.js and React. Comfortable with RAG, knowledge bases, and evals. I keep testing the newer agentic tools as they ship.