Skip to content
● internet trash panda

Hi, I'm Sean.

Co-founder & CTO at B3, building a crypto agent & decentralized inference — previously Coinbase. I sweat the small details in software, and leave the good parts (components, free tools & writeups) out in the open here.

Sean Geng

Components

all →

Writing

all →
A skill that iterates on a plan until it stops improving preview

A skill that iterates on a plan until it stops improving

Treat planning as a search problem: generate a plan, score it against a rubric, critique it, rewrite it, keep the best, and stop when the score plateaus. A downloadable Claude Code skill that hardens any plan — project, code migration, research — and tells you when more iteration is just noise.

agents
skills
claude code
Building a liquid-metal UI kit for React preview

Building a liquid-metal UI kit for React

Argent is Glacé's metal sibling: chrome, gold, and gunmetal surfaces that ripple like mercury. Why CSS couldn't fake this one, how a WebGL shader becomes a component kit, the browser's secret WebGL context cap, and writing my own shader to escape a license.

react
webgl
shaders
Building a liquid-glass UI kit for the web preview

Building a liquid-glass UI kit for the web

Glacé is a small React kit where the glass actually refracts — the edges bend the backdrop like a real lens, not just blur it. Here's the technique (displacement maps, per-element filters, edge profiles), the bugs along the way, and the honest caveats.

css
react
glass
Building agents that live in your messages preview

Building agents that live in your messages

Agents feel best inside the apps you already text in — iMessage, Telegram, WhatsApp, Slack. But iMessage gives you no buttons, and a Slack bot has to juggle many users and orgs at once. Two very different engineering problems, one goal: make the agent feel like a person. How to architect both, with the gotchas.

agents
ai
messaging
The honest guide to LLM model routing preview

The honest guide to LLM model routing

Auto-routing picks a cheap model for easy prompts and a frontier model for hard ones — automatically. The papers claim 85–98% savings. An independent benchmark found a commercial router doing worse than no routing at all. The real number is ~20–25%, and here's exactly why.

ai
llm
routing
Turning WebGL into ASCII, every frame preview

Turning WebGL into ASCII, every frame

A three.js plane rendered to a hidden canvas, then read pixel-by-pixel and rewritten as ASCII characters in a <pre>. The trick is sampling the render at one pixel per glyph. From B3's ai-arena.

three
webgl
shaders