Built with AI
You might think of duiduidui! as the first Chinese language learning app built entirely on agentic engineering and knowledge architecture. AI agents don't just help you learn inside the app; they built the dictionary, scored the curriculum, and wrote much of the code. Here's what that means, and where the humans fit in.
Agents You Can Talk To
The visible layer is the one you meet on day one: 老虎 (lǎohǔ, "tiger"), the AI tutor built into every copy of the app, which knows the whole dictionary and your entire learning history, and can act inside the app on your behalf. And because we believe your AI should be yours to choose, a public MCP server gives Claude, ChatGPT, or Gemini the same tools.
Agents That Built the Dictionary
The 200,000-entry knowledge graph was synthesized by AI agents in an iterative content-synthesis pipeline. Foundational character entries seed example phrases; each phrase becomes a full entry of its own, with pronunciation, translations, and cultural notes; and the process repeats outward from simple characters to complex sentences. The agents aren't improvising: they work from detailed pedagogical prompts that specify how to build compositional progressions, how to write transliterations that reveal the language's internal logic, and how to make cultural notes meaningful rather than mechanical.
Agents That Scored the Curriculum
Every difficulty score in the dictionary was assigned by AI agents applying a human-authored scoring rubric, record by record, in iterative refinement passes. The same goes for the curated overrides that correct the compositional formula where the language itself isn't compositional. The agents supply the detailed language expertise and the patience to score a corpus of this size one entry at a time; the rubric supplies the principles. The full system, formula and all, is on the mathematics page.
Agents That Wrote the Code
The engineering runs the same way. Coding agents build and refine the app, the server, the data pipeline, and this website, working under human architectural direction and review. Agentic engineering isn't a garnish on duiduidui!; it's the method the whole project is built on.
Where the Humans Fit
Every layer of this has a human at the top. Our founder personally oversees the whole pipeline: he wrote the scoring rubrics, drawing on his own experience learning Chinese and a career in computational linguistics, he authored the pedagogical prompts that guide the synthesis agents, and he reviews what the agents produce and decides what ships. The division of labor is deliberate: humans set the direction, define the principles, and exercise taste; agents supply the scale, the consistency, and the detailed language expertise no single person could match across 200,000 entries.
We say this openly for the same reason we publish our mathematics: we'd rather show you how the system works than ask you to take it on faith. The pedagogy the agents execute is one possible pedagogy among many. We designed it, we think it's good, and we keep iterating on it as we learn from how our users actually learn.