Engineer by training, endlessly curious about new tech, and on the front line of every AI wave. At SenseTime, I took an enterprise vision AI product line from 0 to 1 and grew it to hundreds of millions of RMB in revenue. When LLMs arrived, I co-founded an AI eldercare company and proved a vertical agent business model. I still build products myself with AI coding — the best way to judge an idea is to build it.
Ex-SenseTime · Vision AI Business Lead·Co-founder & CPO, Chenghuang Zhiyi·Tsinghua MEM·Trail runner
Hands-on, end-to-end: requirements, design, development and launch. I built each one from scratch with AI coding and use it every day. You only find the real problems by building.
An AI workflow for B2B product teams. It pulls information scattered across meetings, group chats and knowledge bases into one place, then carries it from customer conversations to research reports, PRDs, proposals and quotes — started with one sentence, on one connected dataset.
Core valueTurns senior PMs' know-how into a team-wide system — first for product teams, then for sales and pre-sales — so the front line can deliver a proposal and quote the same day.
13 report types on demand: one sentence picks the type, sub-agents research in parallel, and every conclusion traces back to a source
What's at home, where it is, when it expires — just ask Xiaowuling. Snap a photo to add items; game mechanics keep people coming back to update it.
book-deep-read↗Breaks a book down into 15 verifiable deliverables, so you absorb it in depth and keep knowledge you can draw on anytime.deep-content-extract↗Long videos and podcasts are slow to absorb and hard to cite. It distills them into ideas, stories and frameworks, plus a core-structure diagram.hera-content-radar-skill↗Where you get information matters more than the information itself. It powers Knowledge Radar: curated top sources only, and every conclusion cites one.hera-article-illustrations↗Generates on-brand HERA robot scenes and hand-drawn method diagrams that make articles livelier and easier to follow.Every chapter started with a hard problem: get it working from 0 to 1, then clear the hurdles of standardization, scale and commercialization.
Led a 20-person team: pivoted from digital ads to business insights on an in-house edge-to-cloud stack; standardized it for integrators; went global with SoftBank and other overseas clients — 30+ leading clients in total.
National research project: 42 cameras track curling stones in 3D in real time, showcased at the Olympics, Paralympics and Worlds, with 10+ patents; the full architecture was later reused for an industrial product line.
Integrated eldercare and digital services for upscale care homes, built on "GP + AI": better operations and service efficiency for the home, and chronic-disease care that helps seniors live longer, healthier lives.
Closer to customers, closer to results.
Design how it sells while designing what it does: who pays, why they buy, and why now. Think like the buyer and the user, not just like a feature list.
Skip the engineering romance: a costly 1% accuracy gain may go unnoticed, but ROI won't. Combine models, engineering, rules and metrics to solve the problem.
AI makes building easy, so judging what is good gets harder. Build layered evaluation, dig into bad cases, and never use customers as testers.
Start with one customer type and one problem to test whether it lands and sells. Scale only once it works — don't start big and broad.
How I form a view: take products apart, study competitors, read primary sources, build things myself, and write my thinking down.
We proved the product worked, but not that the business could scale. The problem was not the product — it was four sets of numbers outside it.
Read (in Chinese) →Across two waves — vision AI and LLMs — the same mistake keeps coming back. What customers buy is lower total cost and more certain outcomes.
Read (in Chinese) →A general method, how to take apart AI products, and how to use AI to do the analysis (in Chinese).
The two markets are merging — coding agents are becoming the engine of office agents; as capabilities converge, competition shifts to context, ecosystem and trust.
Read the report (in Chinese) →Revenue is driven by enterprise and coding; Chinese vendors have all gone open-weight; every model vendor now ships agents, and the line between model and application is disappearing.
Read the report (in Chinese) →Chinese companies lead on shipments, cost and profitability; US companies lead on valuation, models and capital. Few robots yet do routine work on production lines.
Read the report (in Chinese) →OpenAI and Google lead in images, Chinese companies in video, specialists in 3D. A model lead lasts only months; the money comes from workflows.
Read the report (in Chinese) →
AI news is never scarce; judgment is. Every day it compresses dozens of AI updates into about ten sourced, verifiable calls.
A one-person research desk: daily calls and conclusions, compounding into a searchable knowledge base.
Turn every action into an achievement you can see — and every day's time into the long-term directions that truly matter. Where your time goes, your growth follows.
Built on Atomic Habits: every action is a vote for who you want to become. Long-term effort becomes visible, so growth compounds.
Outside work, I love marathons and trail running. Running alone on a mountain trail, I often find an unexpected view around the bend; the ground is solid underfoot and my mind goes quiet. Building products is the same: the direction is set, each step is uncertain, but every step counts.
