Frontend developer in Tbilisi, Georgia
I build AI assistants and the interfaces they live in.
React, Next.js, TypeScript and Three.js, with an AI coding agent doing the typing while I decide, review and test. In two weeks that shipped Eliza, a voice assistant that drives a Windows PC. I take fixed-price projects.
34,000
lines of Python, plus a React and Three.js interface
1,370+
automated tests, all passing
270+
commits in the first 16 days
42 ms
to route a request, on a model I fine-tuned
Three things I build, at a fixed price
Each is about a week of work, with the scope and the price agreed in writing before I start.
A Telegram bot or assistant for your product
Your customers or your team talk to it; it reads your data, answers, and asks before it acts.
From $700 · 5 to 7 days
An interactive 3D section for your site
A hero, a product viewer or a particle piece in React Three Fiber that still loads fast on a phone.
From $900 · 5 to 7 days
A live dashboard for your data
One screen that shows what is happening now, built on the numbers you already have.
From $800 · 5 to 7 days
Work
- ElizaA voice assistant that drives a Windows PC: local speech, an agent with tools, a safety gate, a phone bot and a live dashboard.Python · React · Three.js · Electron · AI agents
- Space TourismA 3D space-travel site with interactive WebGL scenes.TypeScript · Three.js · React Three Fiber
- EkkoShopA storefront for tech products with a working cart.Next.js · TypeScript · Tailwind
- A portfolio you can playA retro operating system in the browser, with a terminal and keyboard navigation.TypeScript · React
Notes
All notes2026-10-12
Who makes money in memecoin pools? I read 58,701 real positions
Buyers in these pools lose on average. The people who lend the pool its money are slightly ahead, and the small ones are not.
2026-10-11
My model's 139 favourite memecoins: 137 were the developer buying his own launch
A model trained to pick winners found one template instead. What it looked like, and why no signal warned of the exit.
2026-10-01
A 421M model routes my assistant's requests in 42 ms
Replacing a 9-billion-parameter model's first decision with a small fine-tuned one: ten times faster and more often right.