Consumer quadruped security robot for the home
Defined a product for household inspection and security tasks around mobility, environmental perception, anomaly detection, and family interaction.
Hello, I am Roy.
I build products and conduct long-term research.
Over the past decade, I have moved from fintech, enterprise AI, and data products into consumer software, imaging, and intelligent hardware. The industries look different, but the work I keep returning to is the same: understand what a new technology can solve, find the users and situations worth investing in, and turn that opportunity into a product a team can deliver and a business can validate.
I have worked at Ping An Bank, Baidu, iFLYTEK, ByteDance, and DJI, and I have also started companies. I have built APIs, platforms, and SaaS products as well as consumer AI applications, imaging software, and intelligent hardware. Many of these projects began with little more than a technical opportunity or an ambiguous need, then had to move through research, product definition, team delivery, launch, adoption, and commercial results.
How I build products
I usually begin with three questions:
- What is the user actually trying to accomplish—not which feature are they asking for?
- What can the technology do reliably today, and what is still only a demonstration?
- How should the team be organized so that one validation can become a product that ships repeatedly?
This is why I focus on end-to-end systems. Model quality matters, but it is only one part of a product. Data, tools, interaction, hardware, deployment, operations, and the business model jointly determine what the user ultimately receives.
How the experience connects
In enterprise products, I worked on data platforms, financial risk, business information, OCR/NLP, and intelligent review systems. That work taught me how to turn complex operating problems into reusable product capabilities.
In consumer products and startups, I worked on intelligent editing, AI video, family imaging, and hardware exploration. That work taught me how to begin with a real task and work through experience, growth, and commercialization.
As I moved deeper into imaging and intelligent hardware, more of my attention shifted to the physical world: how software, algorithms, and hardware work as one system; how AI participates in capture, creation, and family life; and how a capability crosses the distance between demo, productization, and delivery.
Products I have worked on
This is not a company-by-company résumé. It is a map of three kinds of product problems I have handled. Hardware projects distinguish between work I led and work I contributed to.
Hardware and the physical world
Making perception, computing, and interaction one coherent product
Led
From product definition through solution coordination
AI compute box
Explored AI NAS and AI Box concepts that bring local compute, private data, model services, and household content management into one device.
Multimodal digital photo frame
Designed voice input with image and video output: users can retrieve target content from an album in natural language and display it directly on the frame.
Contributed
Product and scenario exploration
360° cameras · Panoramic drones · Follow-me cameras
AI models and applications
From foundation models to a creator completing real work
Multimodal models
Contributed to product work for the pre-training and post-training of AI editing and video-generation models.
AIOS and Video Agent
Organized models, tools, and creative workflows so a user can move from intent to a usable video result.
Machine-learning platform
Supported data, training, evaluation, deployment, and team collaboration so isolated experiments could become continuously improving capabilities.
Data products
Turning fragmented data into systems that can be called, understood, and tested
Marketing and risk-model API platform
Profile and feature platform
CDP · BI · A/B testing
What I am researching and building now
AgentMeasure
AgentMeasure is my main public project today. It proposes an open measurement language for the Agent Capability Economy—from Reach, Choice, and Use to Utility and Value—to describe how agents discover, select, invoke, and derive value from software capabilities.
The core question comes before payments or marketplaces: when software consumers become agents, how should we define an opportunity, a choice, a use, and a call that actually creates value?
Research
My long-term research currently centers on four themes: embodied intelligence, multimodal interaction, on-device AI, and agents. I track companies, products, technical paths, markets, and user behavior, and I try to keep facts, inference, and judgment separate as new evidence arrives.
Small tools
I also turn recurring research workflows into small utilities, including User Demand Research (SURE), iRead, and Bilibili Video to Transcript. They solve concrete workflow problems, but they are not the center of my public narrative.
How I research and write
I usually build an industry map first, then move into companies, products, and technology. In writing, I try to separate facts, inference, and position, while preserving the process by which new evidence changes an earlier judgment.
This site is not optimized for daily posting. It is a public workbench: Research tracks the questions I study over time, Essays preserve judgments worth debating, and Projects turn some of those ideas into systems and tools that can be tested.
Contact
If you are working on agents, embodied intelligence, AI hardware, complex zero-to-one products, or the measurement and commercialization of agent-facing software, reach me through the contact page, X / Twitter, or GitHub.