Product Systems Builder · AI orchestration · Full-stack delivery
I turn ambiguous product ideas into shipped systems: product judgment first, AI agents as leverage, full-stack engineering as the delivery surface, and verification as the operating discipline.
我关注的不是“会用 AI 写代码”，而是如何判断什么问题值得做，设计系统如何闭环，组织 AI、工具和人协作，并把结果交付到真实世界。
Personal site · 全智评 · Thyself
I work across frontend, backend, and AI tooling, but I do not treat those as identity boxes. They are surfaces of one job: own the path from problem definition to shipped product.
The durable version of that role is the person who can:
- frame a vague problem into a concrete product constraint
- design the system loop across interface, data, backend, AI, deployment, and feedback
- delegate execution to agents without losing judgment or accountability
- verify outcomes with tests, real-device checks, logs, and written retrospectives
| Project | What it is | System angle |
|---|---|---|
| 全智评 | AI evaluation system for practical experiment and vocational-skill videos | Productized an AI scoring workflow across teacher/student interfaces, upload pipeline, ASR, key-frame extraction, multimodal scoring, asynchronous jobs, storage, and deployment. |
| Thyself | Finite, explainable longform recommendation product | Built around a daily 8-article panel, readable recommendation reasons, Persona modeling, quality gates, LLM reranking, caching, and a browser-extension feedback loop. |
| 触见千年 | Apple Vision Pro cultural-heritage interaction project | Turned the Zenghouyi bells into an immersive, gesture-driven spatial experience with SwiftUI, RealityKit, ARKit, 3D scene loading, and interaction optimization. |
| Agent Workflow | Activity-planning and multi-step agent orchestration experiments | Moved from ad-hoc scheduling toward LangGraph, checkpointing, interrupts, Saga compensation, deterministic constraints, and E2E verification. |
Problem -> Constraints -> Plan -> Build -> Verify -> Reflect -> Reuse
- Product judgment: clarify users, constraints, success signals, and what not to build.
- AI orchestration: use agents for exploration, implementation, review, and repetitive execution, while keeping human judgment in the loop.
- Engineering delivery: connect frontend, backend, storage, async jobs, deployment, and observability into one working system.
- Verification discipline: test the actual symptom, inspect the diff, check the live surface, and avoid claiming success without evidence.
- Knowledge compounding: convert problems, project lessons, and feedback into reusable notes, wiki links, and next actions.
- Product systems that combine interface, data, AI capability, and feedback loops.
- Agent workflows that are inspectable, interruptible, recoverable, and testable.
- Frontend quality under real constraints: mobile, WebView, canvas/WebGL, accessibility, and fallback behavior.
- Practical AI engineering: context design, quality gates, evals, rollback paths, and human review points.
- Build the smallest system that can prove the real behavior.
- Keep AI powerful, but never unaccountable.
- Prefer clear constraints over clever prompts.
- Make every hard-won lesson reusable.