AI at work, from hands-on practice: agents, products, models, hardware and live demos where you can see it for yourself.
Holds the as-built documentation engineer’s position at a construction company: reviews site documents, checks completeness and prepares concealed-work certificates. Asks a person for missing information.
How it works
Finds procurements under 44-FZ and 223-FZ, reads the lot documentation, checks requirements and risks, writes a verdict on each lot.
Answers customers about products, services and company policies. Prices and stock come from internal sources, so answers never drift from what is in the warehouse. Enquiries turn into analytics: what people ask for and could not find. In messengers it understands voice messages and hands complex requests to an operator.
Catalogues and policies stay inside the company. Ask in your own words, get an answer linked to its source document. Runs on a local model.
An AI visualisation and design-development engine for landscaping: a private plot, a residential courtyard, a neighbourhood, the grounds of a commercial property. Interiors run on the same engine with a different input. It grew inside a working landscape studio, which remains its first user and reviewer.
A photo or a plan goes in; a controllable result comes out: a visualisation, its refinement, object placement, a render from a plan and a “from photo A to photo B” document that says what to do, in what order and what to check on site.
How it works
FastAPI · Vue · PostgreSQL · Redis · Docker
A living map of AI tools for construction, landscaping and interiors: what exists for a task, whether it is alive or an abandoned demo, how people use it and what those who tried it say. Every conclusion links to a source; there are no rankings.
How it works
TypeScript · pi agent engine · Alpine.js · Playwright
PDF · DOCX · XLSX · DWG / DXF · OCR
A test question set separates retrieval failures from model errors. The chain: refine context → retrieve → rerank → model; every fragment keeps its page number, so an answer can be checked against the book.
Qdrant · Zvec · GraphRAG · multimodal RAG
With only a few files, the agent reads them directly, without a vector store. That is how the assistant on alesyadesign.ru works.
The choice starts with data, workload and budget: what may go to an external service, what must stay inside, and who will operate the system.
Qwen, Gemma, Llama, GLM, Phi and GPT-OSS are tested for speed, memory and quantisation loss. Qwen-VL and DeepSeek-OCR read drawings, scans and document photos.
Speech to text: Parakeet and GigaAM locally, Groq and Deepgram in the cloud. Tasks reach agents by voice.
Hands-on experience assembling and configuring servers on RTX 3050–5090 GPUs, including modified 48 GB RTX 4090s: hardware, OS, networking, remote access, firewall.
Serving: vLLM · llama.cpp · SGLang · Ollama · LM Studio · Open WebUI
An agent can live on the company’s own server: the model sits next to the data, and only what is explicitly allowed goes out.
Vercel when the platform is enough, a dedicated VPS with nginx when control matters. Certificates, CDN, backups, link previews in messengers.
Metrics and failure alerts; external checks notice when a service stops answering from outside.
Physical machines and VPS: several domains on nginx, self-hosted RustDesk, hosts linked through Tailscale.
Docker · nginx · supervisor · pm2 · fail2ban · Prometheus · VictoriaMetrics · Alertmanager · Grafana · Termius · SOCKS5
Task → architecture → BDD scenarios → code → verification. Claude Code and Codex are in daily use, each in its own role; Droid, OpenCode and cliproxy regularly.
Claude Code · Codex · Droid · OpenCode · cliproxy · Antigravity · Grok · herdr · Orca
Giving one task to different models exposes problems in both the result and the instruction. The rtk router brought no savings on real workflows and was dropped. A custom orchestrator compared Codex, GLM, Sonnet, Haiku, MiniMax and Kimi on the same task; the experiment is complete, and daily work uses existing tools.
Librarian, architect and developer roles; a folder template where an agent runs a project from files and rules rather than scattered chats. Hermes and OpenClaw are deployed and adapted to the work.
When source material cannot leave, generation stays on a local server. The practice covers ComfyUI, FLUX and Qwen-Image.
Clips, promos and social media material: generated with Seedance, Kling and Runway, then edited and scored. For example, vertical clips made from still AI images: the camera flies over a site plan, a person walks up to a finished house.
The material and key points are checked first, then comes the layout. HTML templates and Python turn it into carousels, reports, dashboards and PDFs.
Astro · React islands · Tailwind · Framer Motion · Playwright
A video link → an audio track, with large files split into parts. A channel monitor catches keywords and forwards messages to the right group. Parsers, change notifications, scheduled jobs.
Telethon · yt-dlp · ffmpeg · cron · supervisor
The Green Lecture Hall at the Gardens and People festival, Moscow, August 2026: concept, programme, organisation. Each talk was one specific case, taken apart in detail.
Where AI saves time, where it needs oversight and why one model behaves differently at different steps.
