Leon Liu·柳文浩
Speaker · Open AGI Builders Day · Alibaba × Sentient × Fudan · Shanghai · 2026
Technical developer advocate and hands-on AI builder. I turn real product use into open-source tools, tutorials, talks, and feedback teams can ship.
Then I turn the lessons into demos, tutorials, talks, and product feedback developers can use.
Built Teller during Sentient Arena Cohort 0 and reached 71.5% accuracy at $1.85 per run with MiniMax M2.5. Sentient Arena is backed by Founders Fund, Pantera, and Franklin Templeton.
Teller turns that benchmark work into an open-source grounded-reasoning agent for enterprise document QA.
While preparing an EvoSkill setup tutorial published by Sentient Labs, I reproduced and isolated two Fireworks/OpenCode integration failures, documented a complete MiniMax run, and reported them in issue #48. Fireworks support shipped as a first-class EvoSkill provider in v1.3.0.
First place in the Film & Simulation track at the Worlds in Action World Model Hackathon, hosted at Founders, Inc. Built the backend and end-to-end generation pipeline for turning children’s drawings into explorable 3D worlds.
A verification-first curriculum engine for software tasks, with deterministic generation, fail-closed Docker verification, solver-panel calibration, audit trails, DuckDB lineage, and dataset export. The current local system has been exercised through a 50-task, 250-attempt Fireworks calibration.
A grounded-reasoning toolkit for high-stakes document QA, with a public methodology, competition report, and reusable agent configuration.
On stage, on camera, and inside the issue tracker—the material comes from systems I have actually used.
Speaker at Fudan Science Park in Shanghai on open AI, agent infrastructure, and building across ecosystems.
Event page →
Featured by Sentient Labs discussing agent evaluation, EvoSkill, and what makes China’s open-source AI ecosystem distinctive.
Watch the official X interview →One piece proves the builder, one proves the educator, and one proves the bridge.
How agent orchestration and a filmmaker’s instincts turned a prototype into a full-stack narrative product in five days.
SFT, RLHF, and behavior shaping explained through an ordinary analogy without flattening the technical idea.
A bilingual analysis of the people, models, and assumptions moving between the U.S. and Chinese AI ecosystems.
Film taught me to make complex ideas legible and memorable. Crypto taught me to work in fast-moving developer ecosystems. Today I build agent and evaluation systems, then turn the lessons into demos, tutorials, talks, bug reports, and open-source tools. I work natively in English and Mandarin across the U.S. and China.
Film remains part of the toolkit: visual direction, model fluency, and production judgment across American and Chinese creative-AI systems.