AI Product Builder · Co-Founder & Product Lead, ZenMux.ai
I turn models, tools, protocols, and real production traffic into shippable AI product systems. Engineering and algorithms background, now product — 8 years across LLM infrastructure, agent harnesses, developer tooling, and autonomous-driving perception.
Daily heavy user of Claude Code, Codex, and Cursor — running ~1.5B tokens/month of personal consumption through ZenMux itself.
This repo powers thinkthinking.ai, my personal site.
- Vibe Working — an interactive field guide to extending Vibe Coding across the full product pipeline: moving bottlenecks, OPC and collaborative operating modes, Context as the shared interface, executable artifacts, and super-individual organizations.
- ZenMux.ai — "the LLM insurance company." Unified OpenAI / Anthropic / Gemini protocols, provider routing, automatic failover, fallback models, observability and billing — call insurance and high availability for agent and AI-coding workloads.
- Tbox, Ant Group — product lead for the agent development platform: build, orchestrate, debug, and publish, end to end, for developers.
- AgentOS — co-founder & product lead of an enterprise agent platform. Proposed "Copilot design facing Autopilot," led the end-to-end architecture across a ~30-person team, shipped to 20+ enterprise and government customers.
- Baidu — AI product manager for AI Studio and PaddleX; product owner for PaddleDetection and Paddle3D open source (10k+ stars, GitHub Trending).
- CATL — autonomous-driving perception engineer: 2D/3D detection and tracking shipped on-vehicle, wireless BMS architecture in mass production.
Research — ZenMux Arena
Studies built on real production traffic, not benchmarks:
- Who Are You? — cross-vendor model identity. 27 models, 16 vendors, 10 languages, 29,700 answers.
- Token Economics — the "DeepSeek execution line": how a price anchor rewrites the LLM market.
- Token Deals — real ZenMux billing data: 353B+ tokens, $130k+ saved, 69% average discount.
- Agent Harness — Model + Harness = Agent, tracking ~400 agent papers, projects, and products.
- The DeepSeek Execution Line (2026.06) — what happens when every "Eastern model" is priced like DeepSeek?
- We Spent $1,000 Asking 27 Models "Who Are You?" (2026.06) — 16 vendors, 29,700 answers.
- Agent Product Design (1): Break It to Build It (2023.12) — deconstructing agent products from first principles.
- ZenMux/zenmux-arena — Arena research code and data pipeline
- thinkthinking/skills — Claude Code agent skills
- thinkthinking/cli — personal AI workflow CLI
- PaddleDetection / Paddle3D — former product owner, 10k+ stars