Building
production apps
through vibe coding

In late April I launched the «AI Doomsday» channel — a coordinate system for assessing how close the industry is to the point where AI development slips out of control. The metaphor is borrowed from the Bulletin of the Atomic Scientists, who have run the Doomsday Clock for nuclear threat since 1947. Starting position: 11:55. Every major AI event moves the hand with reasoning — or it doesn't.

The manifesto established the frame. April 2026 compressed into one month what used to stretch over years — releases of GPT-5.5, Grok 5, Claude Mythos. Mythos turned out to be a special case: for the first time in its history, Anthropic refused to release a model publicly, granting access to eleven organizations via Project Glasswing to hunt vulnerabilities. Two weeks later Google answered with the opposite doctrine — a universal Gemini 3.1 Pro plus a fleet of security agents. Two strategies for one threat: one lab closes the model, two others wrap it in control infrastructure.

The Dragunsky fail (April 29) is a textbook example of what happens when AI filtering operates on substrings. Eksmo enabled AI screening of manuscripts for «drug propaganda» under the law that took effect on March 1, 2026. Three weeks later the model flagged writer Denis Dragunsky's surname — because «драг» matched the English «drug». Pushkin, Gogol, Tolstoy, and a Bulgakov biography fell under the same filter. Error type — lexical match without semantic understanding, recall-over-precision optimization at maximum sensitivity.

Legal AI as a product class crossed the point where automation savings are zeroed out by the cost of errors (May 4). In the first months of 2026, U.S. courts imposed 145,000 dollars in fines for hallucinated AI citations in legal filings. Escalation by the quarter: 2,500 in January, 7,500 in March, 30,000 in early April, 110,000 in Oregon on April 4, and Greg Lake's indefinite suspension from practice on April 16 in Nebraska. Meanwhile, 61% of federal judges use AI themselves. The systemic problem is in the deal architecture, not in adaptation speed.

The clock moved backward for the first time on May 5 — from 11:55 to 11:53. Reason: mechanistic interpretability has for the first time crossed from academic niche into engineering practice. MIT Technology Review listed it among the 10 breakthrough technologies of 2026, ICLR 2026 in Rio held a dedicated workshop, and on April 30 the first startup released a public LLM debugging tool. Anthropic's Microscope decomposes the model's activation superposition into interpretable features — you can see what exactly happens before a hallucination or jailbreak. Two-minute shift, not five — because the tool is confined to Anthropic and the industry hasn't matched it yet.

The discount fail on May 6 — an English online store is legally bound to fulfill an 8,000-pound order at 80% off because the chatbot promised it at 5 AM. An hour of flattering questions, gradual coupon escalation from 10% to 80%, a fake code in the order comments. Under UK consumer law, the business is liable for AI promises as for those of a rogue employee. Error type — classic prompt injection via social engineering: the bot was designed with an open scope of competence. Per IEEE S&P 2026, 13 percent of e-commerce sites run chatbots with the same open architecture.

What I take away from the channel's first week. Mythos and Microscope from Anthropic are the only strong movements in the right direction. The rest of the industry keeps shipping the maximum possible. Regulation is fragmented and reactive. Financial penalties for careless AI use in legal-tech already grow faster than savings from AI itself; healthcare is next. Five minutes to midnight — a compromise between «AGI is around the corner» (no, it isn't) and «everything is under control» (no, it isn't).

It all started with chatbots. While setting them up for business tasks, I kept hitting limitations: bots lacked autonomy, context, and decision-making ability. I wanted systems that could act independently, not just answer questions. That's how I dove into automation, neural networks, and eventually arrived at what's now called vibe coding.

For me, vibe coding isn't just "asking AI to write code." It's a chain of interconnected processes: first visualizing the end product, then decomposing it into dozens of small tasks, writing clear instructions for the AI agent, the "magical" code generation process itself, testing, optimization, and finally — launching to production. Each step requires understanding architecture, business logic, and user experience.

My first projects were classic: document workflow automation, marketing analytics collection — tasks I understood well from 15 years in marketing. But complexity grew: an app for a cosmetologist with YClients sync, a lead generation platform with AI scoring, email campaign analytics with GPT-4 Vision. Each project was more ambitious than the last.

The main lesson — practice, practice, and more practice. Don't mindlessly repeat everything you see on social media. It's far more effective to come up with your own project that solves a real problem and bring it to production. Real understanding comes through deployment, debugging, and scaling.

This path is for anyone willing to understand how LLM models and agent systems work. For those who crave new information and are ready to absorb it in large volumes literally every day. A CS degree isn't required — but curiosity and persistence absolutely are.

At first, like everyone, I used the cloud. But as projects grew more complex, server costs kept rising while resources became insufficient. At some point it became obvious: serious AI development needs its own infrastructure. That's how UGREEN DXP4800+ appeared — a NAS with Intel Pentium Gold 8505, 64 GB DDR5, and a 1.8 TB SSD for Docker containers.

The architecture ended up two-tiered. The NAS runs 40+ Docker containers: Claude Code development environment, CosmoDoc, n8n automations, Dify with RAG pipelines, Ollama for local LLMs, Jellyfin, Immich, and the full media stack. An Amsterdam VPS serves as the front: XRay Reality on port 443 disguises VPN traffic as regular HTTPS, while Nginx on :8443 serves web applications. SSH reverse tunnels forward ports from NAS to VPS, providing public access without router port forwarding.

A separate challenge — bypassing restrictions. In Russia, not only Anthropic and OpenAI APIs are blocked, but also Cloudflare proxy, TMDB, api.radarr.video, servarr.com. The solution is an xray-client container with VLESS Reality, routing all blocked traffic through the VPS. Cloudflare DNS only works in DNS-only mode — proxy mode is blocked in Russia. Every new service requires checking: does it need a proxy?

Among the pitfalls: UGOS kernel ACL blocks shared folder access for non-root users, custom resolv.conf breaks Docker DNS, fail2ban on VPS triggers bans from multiple SSH connections. Each problem is experience that saves hours in future projects. Self-hosted isn't just about cost savings — it's full control over data, performance, and architecture.

My main experience with multi-agent systems comes from deploying Dify Platform on self-hosted infrastructure. 11 Docker containers: Flask API server, Next.js frontend, PostgreSQL, Redis, Weaviate for vector search, a sandbox for safe code execution, Celery workers for background tasks. The visual workflow builder lets you chain LLM calls, RAG queries, and custom tools without writing code.

The toolkit includes Claude Agent SDK for programmatic orchestration, MCP servers for connecting external data sources, and Dify itself as a platform for visual AI application design. Each tool fills its niche: SDK for custom logic, MCP for integrations, Dify for rapid prototyping and team collaboration.

The main challenge I faced — context preservation. When a multi-agent system is used by a team, it's critical that each member has access to colleagues' query context and can retrieve it. RAG pipelines with Weaviate partially solve this: documents, conversations, and previous query results are indexed and available to all agents. But a complete solution requires thoughtful memory architecture — short-term, long-term, and episodic.

Where is the technology heading? It's impossible to predict the scale of changes. Perhaps we're on the verge of fully autonomous multi-agent systems that handle the entire development cycle independently — from requirements analysis to deployment — only requesting final approval from the developer to launch the finished product. Agents already write code, test it, and deploy. The only question is when the level of trust will allow removing humans from the loop.