Source, not state.
Before pixels, there is readable source. Every render keeps the reason behind the frame.
HYPIT AI · SHENZHEN · 2026
VIDEO AS SOURCE / 001
Narratage turns narrative into source code — readable by people, executable by agents.
00 / THESIS
A timeline remembers operations, not intent. Narratage records why first, then resolves when, where and how an idea appears. The same narrative source can be inspected, revised and compiled again.
01 / MATERIALS
Before pixels, there is readable source. Every render keeps the reason behind the frame.
Hooks bind intent to narrative anchors. Change one idea without breaking the whole film.
Whiteboard language enters the compiler. What leaves must be visible, editable and ready to render again.
02 / FIELD NOTES
“Video language” on one board; SVML, Hook and LLM diagrams on another. Around them: a shared table, several laptops, and one founder portrait that refused to look like a founder portrait.

The board asks for a high-level video language: developer-readable, editing-free and built to scale.

Intent, raw material, edit, overlay; Hook, encoders and an LLM — the system drawn before it was typeset.

Not a team portrait: one table, several laptops, everyone already inside the work.

Alvin · founder. Fries, iced tea, and a portrait with no attempt to look official.
03—06 / COMPILER JOURNEY
SCROLL DRIVEN / 4 STAGES
WRITE
ANCHOR
GRAPH
COMPILE
07 / PEOPLE
Average age 20. As the founder, I’m the one dragging it up.
Seven people, from a 16-year-old high-school builder to multimodal researchers, infrastructure engineers, product designers and serial founders. No departmental relay race: language, product, content and distribution stay around the same source.
Founder
Peking University, class of 2025
PKU class of 2025 · serial AI founder · a side project at 1.3k GitHub stars without a launch or paid distribution.
Multimodal research · AI content
Tsinghua University, CS, class of 2022
Tsinghua CS · KEG and Zhipu multimodal research · built a 300k-follower AI persona in two months.
Backend · infrastructure
South China University of Technology
Tencent and ByteDance backend · mini-program foundations, test infrastructure and service reliability · TikTok shop operator.
Product experience · design
M.S. Information Science, Cornell · B.A. Design & Psychology, Pratt
Cornell Information Science · Pratt design and psychology · AI product design, jewelry factories and hands-on sales.
AI-native design & frontend
Peking University, School of Physics, class of 2022
PKU Physics · PHYBench co-first author · Clab core developer · long-time open-source contributor.
Intern · AI video & agentic workflow
Independent developer, 18
18 · six years running creator channels · seven million-view videos · turns production into agentic workflows.
Marketing
High school student, 16
16 · built a web product used 100k+ times in one week · now preparing and running Hypit marketing.