Full-stack developer. I design, build and ship AI agents into real products.
TypeScript, React, Next.js, Astro, Node.js, Cloudflare Workers, Anthropic API, MCP
The model never computes a number.
An AI finance app where every number comes from SQL: ~97% categorisation accuracy at ~80% lower cost.
The agent must look at what it rendered.
Recolours a Lottie animation into any palette or theme, by hand, by script or by an agent that checks its result by rendering it.
Personal projects & agents
LiveWeb appMCP
An AI finance app where every number comes from SQL: ~97% categorisation accuracy at ~80% lower cost.
A multi-user finance app that syncs Monobank and CSV statements, categorises spending and answers questions about your money. Each user's data lives in its own Durable Object. A small judgment model and Claude share categorisation in a measured cascade. The same numbers are exposed to Claude over an MCP server with built-in OAuth 2.1.
TypeScript · React · RTK Query · Cloudflare Workers · Durable Objects (SQLite) · D1 · Hono · Anthropic API · TypeSafe Jev
LiveMCPOpen source
Recolours a Lottie animation into any palette or theme, by hand, by script or by an agent that checks its result by rendering it.
A browser editor, a CLI, an MCP server and a live editor–agent bridge that find every colour in a Lottie file, remap it and render the result to check it. All four are thin shells over one core package, published to npm, so a hand edit, a script and an agent come out identical.
TypeScript · Next.js · React · Zustand · lottie-web · Tailwind CSS · shadcn/ui · MCP SDK · Anthropic SDK · GitHub Actions
v2 in progressBotRAG
A group-chat AI bot where code decides when the model runs: an estimated $3-6 a month instead of $75-150.
A serverless Telegram bot for a friends' group chat. It logs messages, summarises on demand and answers as a persona with semantic memory. An LLM intent classifier routes requests, and cheap deterministic gates decide whether a model is called at all.
TypeScript · Cloudflare Workers · grammY · D1 · xAI Grok · Workers AI · Vectorize · R2
LiveToolAutomation
A job-search tool where code makes every decision and the model only reads new text: cost down from ~$15 to an estimated $1-2 a month.
A personal tool that collects vacancies and company catalogues, filters them with plain code and calls a cheap model only on new text. The score is deterministic. I pick every company and send every letter myself; the tool only drafts a first paragraph, and code checks it before I see it.
TypeScript · Hono · Drizzle ORM · Cloudflare Workers · D1 · Anthropic API · Workers AI · React · Mantine · Zod
LiveAgentLLM
The AI chat on this site. It answers only from my real content, for under a cent a question.
A streaming chat that knows my projects, experience and how I work. The whole knowledge base is built from this site's content at build time and sits in a cached system prompt, with rate limits, a daily budget and a golden-question eval set.
Next.js · TypeScript · Anthropic API · Upstash Redis · Zod
3+ yrs · remote
Full-Stack Developer
10/2023 – Present · Contract · Remote
Next.js · React · TypeScript · Astro · Node.js · Strapi · PostgreSQL · Firebase · Stripe · Cloudflare Workers
Front-End Developer · client: Bold, UK
06/2024 – 05/2026 · Agency · Remote
HTML · PostCSS · JavaScript · Figma
Front-End Developer
03/2023 – 10/2023 · Freelance · Remote
JavaScript · HTML · CSS · WordPress
Via Sollas · Metamorfosi · Freelance
Kyiv, Ukraine
Hi, I'm Vitaliy, or just Italik. I'm a full-stack developer from Kyiv who's been shipping for the web for 3+ years, taking products from a Figma file all the way to production.
Now I focus on AI: I build agents from scratch, wire LLMs into real products and keep them honest about the data. In Money Track the model never computes a number.
I enjoy owning a feature from idea to something that actually ships. I work remotely from Kyiv, alongside designers and, where there are any, the client's own developers, and I'm happy to own the whole stack or slot into an existing team.
Next, I want to build AI products with a team, as an AI engineer or full-stack developer. Full-time, contract or part-time all work for me.
Live demo
It knows my projects, experience and how I work. Ask it anything a recruiter would.
Site content and my notes are loaded into one cached prompt. The model answers only from them and cites the files it used.
Italik.aiExample answer
It keeps the model away from arithmetic. Every number comes from one canonical SQL layer over D1, and each user's data is isolated in its own Durable Object. The LLM advisor only receives those query results as grounded context and explains them; it never computes totals or writes to the data itself.
Work in progress
BuildingAgentTool
A hand-written agent harness in TypeScript on the raw Anthropic SDK, with no agent framework. The model navigates a document through a section-reading tool, writes quiz questions as validated tool input, and code grades the answers.
2026-09 · CLI works; building the web version: Next.js, Hono, sessions in SQLite
Repov2 in progressBotRAG
A ground-up rewrite: function calling instead of a hardcoded router, tests, and smarter memory recall.
2026 · v1 runs in a live chat; rewrite underway
RepoNext upAgent
Not startedAn agent-mentor that builds a personalised learning course.
Next · Starts after Doc Quiz
No repo yetBy area
