AI PRODUCT · AGENTS · EVALUATION
维怀·weihuai
Building and researching practical AI systems, with a focus on agent architectures, real-task evaluation, and context engineering.
Previously, I worked on enterprise AI delivery at Zhipu AI and model iteration strategies for the Doubao family at ByteDance. Today, I explore the boundaries of human-AI collaboration as an independent builder and writer.
This site collects my open-source experiments, product essays, and field notes—written for people who want to clarify their problems and build AI tools that actually hold up in the wild.
Background2019 — 2026
ByteDance
2024.11 — 2026.04
AI Strategy Product · AML
Grounded Doubao model iterations (including Coding models) in external users' real tasks and failure traces rather than static leaderboards. Co-built the Volcengine Ark support agent from scratch, pioneering a hierarchy-navigation retrieval approach.
Zhipu AI
2023.10 — 2024.11
AI Product & Enterprise Delivery
Worked on the front lines of 40+ enterprise deployments across agents, customer support, retrieval, and content moderation. Focused on bridging the gap between impressive demos and production reality—aligning controllable behavior with shared acceptance criteria.
Earlier Product Practice
2019 — 2023
FiberHome Zhongzhi · Banma · Leigod
Before transitioning to foundation models, I designed and shipped complex systems across smart cities, in-vehicle infotainment, and consumer software. That grounding taught me to approach AI through actual workflows, end users, and accountable outcomes.
Selected practice
All practiceVolcengine Ark support agent
Moving beyond fragmented chunk matching: teaching an agent to navigate authoritative documentation like a human reader.
Iteration strategy for Doubao Coding models
When static benchmarks plateau, how do you turn anonymous real-task testing and failure traces into model improvements?
Enterprise AI design and delivery
Turning "what the model can do" into explainable, mutually accepted engineering outcomes.
Writing
All writingA Ruler for Measuring the “AI Feel” in Writing
Instead of guessing whether a human or AI wrote a piece, I started with whether a reader wants to keep reading—using 435 pairwise comparisons to find interpretable linguistic signals.
The Demo Fooled Everyone: AI Costs Add Up Quadratically
Unpacking the compute cost curve of AI agents, and why any evaluation that ignores cost and non-AI alternatives is incomplete.
Before the First Line of Code: Thinking Through an AI Application from Scratch
A mental model for framing the problem, context boundaries, and verification loops before touching an API or prompt.
Independent explorations
ACRON
An open-source agent exploring context signal-to-noise ratio in long-running collaboration by decoupling intent from execution traces.
Let’s talk
If you’re building AI products, designing evaluations, or figuring out how agents hold up in the wild, I’d love to exchange notes.