I write and build independent projects under the name Weihuai.
Much of my career has revolved around a single question: When a new technology meets a messy real-world environment, how do we know if it actually works?
From designing smart-city and in-vehicle software early on, to leading enterprise AI delivery for 40+ clients at Zhipu AI, to shaping Doubao model iteration strategies and support agents at ByteDance—I’ve learned that the bottleneck in applied AI is rarely raw model capability alone. More often, it lies in how precisely the problem is framed, the signal-to-noise ratio of the context, and whether evaluation metrics reflect real human experience.
Today, this site serves as my digital workshop—a place to document field-tested methods, open-source experiments, and ongoing essays.
How experience shapes my work
Whenever I approach an AI product, I start with two questions: What exact problem are we trying to solve, and what evidence will convince us it has been solved?
During my time at ByteDance, a stalled content-moderation project brought me face to face with how differently teams define "good performance." Engineers looked at offline accuracy; the client watched day-to-day shifts in online detection rates for specific violation categories. To break the deadlock, I rebuilt the process around stratified sampling, blind dual-annotation, and disagreement resolution—turning abstract debates into shared criteria.
My work at ByteDance also expanded beyond individual clients to foundation models and developer ecosystems. Static benchmarks were easy to saturate, yet users still struggled with real coding and troubleshooting workflows. Whether designing anonymous real-task studies for Doubao Coding models or building the Volcengine Ark support agent from 5,000 raw tickets, I focused on connecting execution traces directly to model and architecture improvements.
Writing and building open-source tools are how I structure what I learn. When sharing a method, I care as much about its preconditions, trade-offs, and failure modes as its success metrics. Good frameworks don't promise silver bullets; they help builders see boundaries clearly.
Outside of commercial work, I build independent tools to test ideas that won't leave me alone. I believe working with AI should demand less manual babysitting of tools and context, leaving human attention for the problem itself. This site keeps both the working prototypes and the questions that remain open.
Before moving into foundation models, I worked on products at FiberHome Zhongzhi, Banma, and Leigod, shipping software across smart cities, automotive cockpits, and consumer utilities. Navigating those complex, multi-stakeholder systems continues to ground how I view AI engineering today.
Three principles I keep returning to
Align on acceptance criteria before optimizing
Without a shared definition of success, faster iteration just means drifting faster. Breaking vague ambitions into observable, attributable samples is where engineering begins.
Evaluate through real tasks and failure traces
Scores on static test sets hide real-world fragility. A meaningful evaluation must capture who is doing what, how it feels in practice, and the exact step where an execution trace breaks down.
Curate context with discipline
Whether retrieving documentation or selecting from hundreds of APIs, stuffing the context window is rarely the answer. Every layer of structure and every token introduced should justify its role in completing the task at hand.