
sapgun
Building local-first AI and Web3 tools in Busan. Currently focused on Vespera.
More tools were supposed to mean less work. Somehow, I ended up managing the tools.
Working on Vespera — a local-first agent runtime focused on owned context, human approval, and recoverable task state.
Digging into agent permissions, durable context, and what happens after a task stops.
Also bringing people together around Web3 and AI in Busan through BUKMO.
Updated October 2026
Vespera
A local-first agent runtime built around user-owned context, human approval, and recoverable task state. Built with Tauri, React, and Python; the core is private and the website is public.
More agents. Preferably fewer loose ends.
vespera-webTrustFi
A Web3 gateway exploring how to make risks understandable before users sign or grant permissions.
Connecting a wallet is easy. Knowing what you just allowed should be, too.
TrustFiAI Systems Bottleneck
A practical guide to reasoning about AI system bottlenecks, from symptoms to the constraints underneath.
Before swapping the model, find the bottleneck.
book
Useful detours, tools, and experiments. Some rabbit holes come with source code.
- KeyGlow
A visual keyboard controller for Windows. Includes a Cat Lock, because keyboards have other users.
- Media Grabber
A small Windows interface for yt-dlp and FFmpeg. Paste a URL; spare yourself the flags.
- Vespera Motion Lab
Canvas experiments for context, approval, and continuity. Making abstract ideas move.
Notes from building, reading, and occasionally changing my mind.
“Nice site” is welcome. A reproducible bug is even better.
First here? Leave a footprint.
Built in Busan.
