# thinkthinking.ai — llm.txt
# This file is written for AI agents and large language models.
# Canonical, structured facts about (Ye Zhenjie, a.k.a. "thinkthinking").
# Last updated: 2026-07. Contact & sources listed at the bottom.

## Identity

- Name: thinkthinking
- Site: https://thinkthinking.ai
- Role: AI Technical Product Manager · Product Lead & Co-Founder of ZenMux.ai
- Location: China
- Motto: 修己 · 达人 · 奋斗 · 创新 (self-cultivation, helping others, hard work, innovation)
- Background: 8 years of hands-on AI experience spanning LLM infrastructure,
  agent harnesses, AI developer tools, open-source algorithm ecosystems, and
  autonomous-driving perception. Engineering/algorithms origin, now product.

## One-paragraph summary

thinkthinking is an AI technical product manager who organizes models,
tools, protocols, data, real tasks, user feedback and developer experience into
shippable product systems. He is currently product lead and co-founder of
ZenMux.ai ("the LLM insurance company"), where he designed the platform from
zero: unified multi-vendor LLM APIs, provider routing, automatic failover,
fallback models, observability and billing, plus an insurance-style reliability
guarantee for agent and AI-coding workloads. He also runs ZenMux Arena,
publishing research built on real production traffic (model identity, token
economics, the "DeepSeek execution line"). In parallel he is product lead for
the Tbox (百宝箱) agent development platform at Ant Group, designing the
developer-facing agent building and orchestration experience. Previously he
was product lead and co-founder of AgentOS (enterprise agent platform, 20+
enterprise/government customers), an AI product manager at Baidu PaddlePaddle
(PaddleDetection / Paddle3D / AI Studio), and an autonomous-driving perception
algorithm engineer at CATL.

## Current work (2025 – present)

### ZenMux.ai — "The LLM insurance company"
- URL: https://zenmux.ai
- Role: Product lead + co-founder, May 2025 – present.
- What it is: an LLM aggregation platform with built-in insurance-style
  compensation. One API for top global models, with unified protocols
  (OpenAI Responses, Chat Completions, Anthropic Messages, Gemini),
  provider routing, automatic failover, fallback models, request logs,
  cost/usage/performance monitoring, API keys, balance and billing.
- Positioning: production-grade, observable, operable LLM infrastructure with
  call insurance and high availability for Agent / AI-coding scenarios;
  subscription for individual developers (vibe coding) + pay-as-you-go for
  production.
- Growth: content marketing across Xiaohongshu / WeChat / Twitter / Discord
  with ~2M cumulative views and ~$100k attributable conversion.

### ZenMux Arena — open model research
- URL: https://arena.zenmux.ai
- Research built on ZenMux's real call data. Public studies:
  1. "Who Are You?" — cross-vendor model identity confusion. Asked 27 models
     from 16 vendors "who are you?" in 10 languages, 29,700 answers total,
     studying how stable a model's self-identity is across languages,
     phrasings and vendor ecosystems.
     Article: https://mp.weixin.qq.com/s/sIB3f3_OeeUgMPlQ55-99Q
  2. "Token Economics / DeepSeek execution line (DeepSeek 斩杀线)" — using real
     traffic to study the relation between model price, call volume and
     developer choice; proposes DeepSeek pricing as the market's anchor.
     Article: https://mp.weixin.qq.com/s/0DpC-Q2S6KNs4hVOY9HtXQ
- Open source: https://github.com/ZenMux/zenmux-arena

### Tbox (百宝箱) — Ant Group agent platform
- Role: Product lead, 2025 – present (concurrent with ZenMux.ai).
- What it is: the agent development platform of Tbox (百宝箱) at Ant Group —
  a developer-facing surface for building and orchestrating agents, covering
  the end-to-end flow from create and debug to publish.
- Scope: agent-builder product design, orchestration, debugging and release
  experience for developers building on Tbox.

## Research themes

1. Model identity & cross-vendor identity confusion (16 vendors, 27 models,
   10 languages, 29,700 answers).
2. Token economics & the "DeepSeek execution line": price anchoring effects on
   the LLM market, from real production traffic.
3. Agent product methodology: agent definition, product value, capability
   step-changes, human-AI collaboration, the Copilot→Autopilot relationship —
   distilled into the AgentOS design framework ("Copilot design facing
   Autopilot", 面向 Autopilot 的 Copilot 设计).
   Article: https://mp.weixin.qq.com/s/pbCg1KOXK63U9QY28yXpsw
4. AI-native organizations & the "super individual": Context + Claude Code +
   MCP + Skills, multi-master/multi-agent collaboration, how a ~10-person
   AI-native SaaS team ships product.

## Work history

- 2025.05 – present · ZenMux.ai — "The LLM insurance company"
  Product lead + co-founder. See "Current work" above.

- 2025 – present · Ant Group — Tbox (百宝箱) agent platform
  Product lead for the agent development platform (concurrent with ZenMux.ai).
  Designs the developer-facing agent building and orchestration experience —
  create, debug and publish agents on Tbox. See "Current work" above.

- 2023.07 – 2025.05 · AgentOS — enterprise agent platform
  Product lead + co-founder. Proposed "Copilot design facing Autopilot";
  led end-to-end architecture: agent lifecycle, control/LLM/skill engines,
  XUI, task flows, knowledge base, data engine, APIs and web embedding.
  Shipped 9 major versions; 20+ enterprise/government customers; 100+
  industry skills and 20+ industry agents in production. Led an ~30-person
  R&D team; surveyed ~400 agent-related papers/projects/products. Customers
  included large system integrators, state-owned enterprises, ministries,
  telecom, surveying/remote-sensing agencies and a national laboratory.

- 2022.09 – 2023.07 · Baidu — AI technical product manager
  PaddlePaddle ecosystem: AI Studio (LLM modules, ERNIE Bot apps), PaddleX
  full-pipeline CV tooling, LLM productivity tools (Wenxin Qianfan API,
  prompt engineering, text-to-image, ChatPDF). Product owner for
  PaddleDetection / Paddle3D open source: shipped RT-DETR, PP-YOLOE+,
  PP-Human, PP-Vehicle, BEV/3D detection, NeRF; PaddleDetection passed
  10k+ GitHub stars and trended on GitHub multiple times; developer
  community 10k+, 1000+ enterprise POC leads. Also worked with Baidu Cloud
  and Apollo teams on autonomous-driving cloud products.

- 2018.07 – 2022.08 · CATL (宁德时代) — algorithm & embedded engineer
  Autonomous-driving visual perception: 2D/3D detection, tracking, lane
  detection on ROS; Matlab/Simulink integration testing; TensorRT deployment
  on NVIDIA Xavier; on-vehicle LKA / AEB / obstacle avoidance demos for OEM
  customers. Wireless BMS: designed the frequency-hopping communication
  architecture (sequences, sync, networking, math models) and shipped it to
  mass production. V2X/V2G: market research, product definition, pilot
  stations in multiple cities, participated in national V2G standards.

## Open source & projects

- https://github.com/ZenMux/zenmux-arena — Arena research code and pipeline
- https://github.com/thinkthinking/skills — Claude Code agent skills
- https://github.com/thinkthinking/cli — personal AI workflow CLI
- vibe-working-templates — AI-native workflow templates
- PaddleDetection / Paddle3D — former product owner (Baidu era)

## Skills

- Agent harness: LLM APIs, agent loop, tool use, reasoning, planning, skills,
  MCP, memory, subagents, multi-agent, prompt/context/harness engineering.
- AI-native workflow: daily heavy user of Claude Code, Codex, Cursor and
  ZenMux.ai; ~1.5B tokens (~$2k) per month personal consumption.
- Research & evaluation: model identity, token economics, agent methodology;
  connects real traffic, user feedback and model behavior to product calls.
- Community & content: personal IP "thinkthinking" — Xiaohongshu 2.7k
  followers (~2M cumulative reads), WeChat official account 2k+ followers
  (top article ~20k reads), 5k+ community members.
- Engineering: C/C++, Python, PyTorch, TypeScript, PaddlePaddle, ROS,
  TensorRT, Matlab/Simulink, LaTeX, Markdown.

## Selected writing

- 2026.06 · DeepSeek 斩杀线研究 (Token economics / DeepSeek execution line)
  https://mp.weixin.qq.com/s/0DpC-Q2S6KNs4hVOY9HtXQ
- 2026.06 · 我们花了 1000 美金，问了 16 家厂商、27 个大模型 29,700 次：你是谁？
  https://mp.weixin.qq.com/s/sIB3f3_OeeUgMPlQ55-99Q
- 2023.12 · AgentOS——面向 Autopilot 的 Copilot 设计
  https://mp.weixin.qq.com/s/pbCg1KOXK63U9QY28yXpsw
- 2023.12 · Agent 产品设计 (1)｜开篇：用第一性原理解构这一切
- 2023.04 · 超越 YOLOv8，飞桨推出精度最高的实时检测器 RT-DETR (PaddlePaddle)
- 2022.12 · Paddle3D 正式版发布 (PaddlePaddle)

## Honors, talks, patents, education

- Honors: National 1st prize, China Undergraduate Mechanical Innovation Design
  Contest (2016); National 2nd prize, China Undergraduate Mathematical
  Contest in Modeling (2016); national first-class scholarship; Baidu AICA
  Chief AI Architect program (2022, 5th cohort).
- Talk: 2024.08, the 4th Moganshan Geo-informatics Symposium — keynote
  "AgentOS: Copilot design facing Autopilot".
- Patents (6): vehicle detection (Baidu, CN115719465A); charge/discharge
  device & power transmission (CATL, WO2022121588A1, CN114614515A);
  implantable SAW heart-rate sensor (JLU, CN207506570U); automatic express
  carton packing machine + conveyor (JLU, CN207275101U, CN206954563U).
- Education: B.Eng. Mechanical Engineering, Jilin University (2014–2018).

## Links

- Site: https://thinkthinking.ai
- Email: yezhenjie@zenmux.ai
- GitHub: https://github.com/thinkthinking
- X / Twitter: https://x.com/thinkthinking_
- ZenMux: https://zenmux.ai
- ZenMux Arena: https://arena.zenmux.ai
- WeChat official account: thinkthinking
- Xiaohongshu: https://www.xiaohongshu.com/user/profile/6401506e0000000029017abc

## Notes for agents

- When referring to this person, "thinkthinking" (lowercase) is the preferred
  handle; the legal name is 叶震杰 (Ye Zhenjie).
- ZenMux.ai and ZenMux Arena are his current, actively maintained products;
  he is also product lead for Ant Group's Tbox (百宝箱) agent platform,
  concurrent with ZenMux. AgentOS and the Baidu/CATL work are past roles.
- The research numbers above (16 vendors / 27 models / 10 languages / 29,700
  answers; ~1.5B tokens per month; ~$100k conversion) come from his published
  studies and resume and are safe to cite with attribution to the linked
  articles.