Development & AI.
Who builds it, what's already in production, how AI actually works in these projects, what I check by hand, and where real engineering doesn't reduce to prompts.
Used to run a 40-person IT company: credit-card processing, a search engine, online games, startups. Now I'm one. All 40 people became dead weight — meetings, arguments, weekends gone, constant approval of trivial decisions.
Since 2024 — solo. Claude Code and four AI agents replace the whole mid-tier team — juniors, mids, PMs, sometimes seniors on isolated tasks. I stayed as architect and tech lead: ask questions, verify results, make decisions, debug production incidents. The machine writes the code.
Not theory, not marketing. Eleven production projects confirm it. Atlas, a home-delivery platform, delivered 1,035 orders to real customers in September. Jusmila in Lithuania ships motorcycle parts daily. GoraGen generates AI video at goragen.com right now. BRUNO Tbilisi and WedInGeorgia run without downtime. Aperwind syncs a Thailand property catalogue every morning without losing a single URL.
- name
- Gora (George)
- role
- Web developer, architect, tech lead
- format
- Remote, clients worldwide · solo since 2024
- experience
- 12+ years · from junior to running a 40-person IT company · now solo
- production projects
- Eleven: Atlas (home delivery), Jusmila (Lithuania), Aperwind (Thailand), GoraGen (SaaS), BRUNO Tbilisi, WedInGeorgia, Phone farm, GoraBot, VPN, Auto-reply, SEO
- telegram messages
- 31 000+ in Telegram-based projects
- regions
- USA · Lithuania · Thailand · CIS and EU
- languages
- Russian (native), English
- rate
- from $300 (landing) → platform from $5 000
A concrete list. Not «we use AI» — this is what's actually delegated to Claude Code and AI agents without my manual touch on each iteration.
- Next.js / NestJS boilerplate — scaffolding, layouts, routes
- Unit tests against existing code — Claude Code writes, I review
- Translating essays into EN / LT — Sonnet 4.6, native review pending
- Image alt-text, OG and Twitter meta strings
- Generic components per design system (TOC-row, kickers, section frames)
- Documentation drafts, READMEs, commit messages
- MDX frontmatter, JSON-LD from existing data
The machine doesn't touch the following without an explicit instruction — or doesn't touch them at all. These are trust boundaries where automation costs more than control.
- Architecture decisions: schema, multi-tenancy, module boundaries
- Production deploys: systemctl, nginx reload, certbot
- Incidents in production, root-cause analysis, hotfix pipelines
- Secrets: .env files, API keys, OAuth, JWT secrets — never enter the LLM context
- Client communication, timeline estimates, pricing negotiations
- Every git push to production branches and force-push
- Design decisions: component IA, brand DNA, typography
- Security: rate limits, auth boundaries, NDA scope
Tasks where the LLM agent is a junior with twelve years of doc scraping — and no production intuition. Here I think and decide, AI assists.
- Multi-tenant isolation: row-level security, tenant-aware queries, audit trails
- Real-time scaling: Socket.IO + Redis adapter, sticky sessions, presence tracking
- GPU orchestration on Vast.ai (Plan-100 control plane) — worker start/stop tied to queue depth
- DPI-evasion VPN: WireGuard + Hysteria2 + sing-box, fake-CDN masquerade
- Telegram WebApp HMAC validation (mini-apps), session coherence
- RAG pipeline: chunking strategy, embedding model choice, reranking, tuning distance < 0.4
- Core Web Vitals: real TBT/LCP/CLS optimization, font preload, route-level cache
- pgvector + HNSW: index tuning, ef_search, recall-vs-latency trade-offs
- web
- Next.js 15 · React 19 · Vite · Astro · Remix · Tailwind v4 · TypeScript
- backend
- NestJS · TypeORM · FastAPI · psycopg v3 · Socket.IO + Redis adapter · Caddy · nginx
- data
- PostgreSQL 16 · Redis · ChromaDB (vectors) · pgvector · Cloudflare R2 / S3
- AI / LLM
- Claude API + prompt-caching · Gemini · Anthropic Claude Code · RAG · voyage-3 embeddings · MCP
- bots
- Python aiogram 3 · TypeScript telegraf · Telegram WebApp HMAC · Twilio A2P · WhatsApp Business API
- infra
- Docker compose · GitHub Actions → GHCR · Let's Encrypt · Cloudflare · Vast.ai (GPU) · Sentry
- SEO
- 15 SEO skills via Claude · Core Web Vitals · schema.org / JSON-LD · llms.txt · AI Overviews · FAQPage rich snippets
- 01 first message
Telegram, in your own words. I don't need a brief or technical terms — I care more about understanding the task than its formal label.
- 02 reading between the lines
Within 1–2 days I send a timeline-and-budget estimate. If it's not my case — I say so in the first reply, no runaround.
- 03 lock-in
Price and timeline in writing. 50% upfront. After that things change only by mutual agreement, and rarely.
- 04 work
Interim results at my own pace. AI agents write code, I hold the architecture and verify production-critical paths. At delivery — handover with documentation and access. First month of support included.
- Projects aimed at minors — no exceptions.
- Financial schemes against retail individuals.
- Anything built to harm identifiable people.
- Anyone unwilling to sign an NDA.
Everything else — let's talk.