Context Engineer
Kirill Korbut
I ship real, production software by directing AI end-to-end — context, memory, and orchestration, not autocomplete. Every project below, including this site, was built that way.
About
I build by directing, not typing line by line.
I build complete products — solo — by orchestrating Claude across every stage of the work: research, planning, multi-file implementation, and verification. That's not a shortcut; it's a real, deep skill — knowing what context an agent needs, how to structure it, when to plan before executing, and how to check results instead of trusting them blindly.
Six shipped products back that up: a commerce platform, a casino game provider with its own remote gaming server, a crypto-funded Telegram Mini App casino backend, an AI assistant running inside a live Telegram Business account, a payment reconciliation service, and a handful of polished client sites. Different domains — web apps, bots, game engines — same process underneath.
Selected Work — 01 / 06
Vitrea
Digital Goods Store Template
A full-stack commerce platform template, built solo as a reusable example rather than a single fixed store: Next.js 16 + Prisma/PostgreSQL storefront selling licenses, keys, subscriptions and templates in 10 languages, with full RTL support for Arabic. Every key fulfillment runs inside a database transaction with a claim-and-retry guard — a digital key can never be sold twice, even under concurrent checkouts. Two-tier admin (Owner/Staff) with masked customer emails, CSV export, and a built-in, cookieless analytics dashboard.
View live ↗
Selected Work — 02 / 06
Casino Games & RGS
Game Provider
A self-built casino game provider running five live games — two slots (Zeus vs Hades, Book of Awakening), European roulette, crash, and baccarat — on one shared architecture: a headless, seeded math engine verified with a Monte Carlo simulator over millions of rounds, served through a signed, SQLite-backed remote gaming server. Two of the five run on fully custom Three.js 3D tables, wired end-to-end to the live platform, not a standalone demo. A sixth game, blackjack, shares the same engine but is still mid-integration on the client side.
Watch the 3D table deal ↗Selected Work — 03 / 06
Casino Platform
Private Telegram Mini App Casino
Backend for a closed, TON-funded Telegram Mini App casino (a private community, not a publicly licensed operator), hardened by a 7-detector fraud engine — RTP exploitation, net-extraction, multi-accounting, bonus abuse, ledger-integrity breaks — that can auto-freeze withdrawals platform-wide. A four-mechanic bonus system, a streamer referral program paid on net GGR, live-scored tournaments, and genuine Telegram Business API gift redemption round out the wallet layer, all under row-level locking with idempotency keys on every external action.
Selected Work — 04 / 06
LeadCopilot
AI Business Assistant
An AI assistant that runs inside a real Telegram Business account, drafting replies to rental clients and sending only on manual approval. The system prompt is genuinely hand-tuned — ~30k characters of domain knowledge and few-shot dialogue transcribed from the owner's real chat voice. An 8-stage CRM pipeline only ever advances forward; the terminal stages are withheld from the model and require an explicit owner tap, so a bad model call can never silently close a live lead.
Selected Work — 05 / 06
Payment Reconciliation
Webhook & Ledger Service
A self-built service that ingests Revolut transaction alerts — mirrored from the phone's own notifications, since Revolut doesn't expose merchant webhooks at this account tier — and turns them into a real transaction pipeline: content-fingerprint deduplication standing in for a proper transaction ID, and a cents-based routing convention that maps a payment straight to the right seller and order.
Selected Work — 06 / 06
Static Sites
Personal & Client, Zero-Maintenance
A set of zero-build-step static sites: my own 8-language, RTL-aware personal/business site — hand-built glass UI, a full legal public-offer page with a scroll-tracked table of contents — plus several lightweight client petition and donation pages with live signature counters.
Process
How it's built
Every project starts with a written brief — a CLAUDE.md — and a persistent memory of decisions, constraints, and past mistakes that carries across sessions, so context isn't rebuilt from scratch each time. Non-trivial changes get a plan before any code is touched, and research fans out across parallel agents when the scope is broad. The result is software that's actually finished: the process is built to catch gaps before they ship, not after.
Brief
A CLAUDE.md documents the goal, constraints, and decisions for each project before any code is written.
Memory
A persistent, cross-session memory carries context and open items forward — nothing gets rebuilt from scratch.
Plan
Non-trivial changes get a written plan and sign-off before implementation starts.
Verify
Claims get checked against the real code — simulations run, numbers measured, not assumed.
Skills
Two tracks, one workflow.
AI Orchestration
- Context engineering
- CLAUDE.md authoring
- Persistent memory design
- Multi-agent workflows
- Planning before execution
- Verification-first process
Technical
- Next.js / React / TypeScript
- Node.js & Python
- PostgreSQL / Prisma
- PixiJS / Three.js
- Telegram Bot & Business API
Contact
Let's talk.
Open to full-time roles and freelance work — especially anything involving AI-directed development, context engineering, or shipping real products fast.