Vladimir Krutov

Full-stack developer & AI engineer
AI agents · Self-hosted LLMs · Angular · NestJS · Python
Since 2019building commercial web products
3 yearsrunning AI models on my own hardware
72 GBof GPU memory in a server I built myself
97–100%“human” score for texts from my AI pipeline

About

I build web products end to end and the AI systems that take over routine work around them: content, SEO, page layout, reporting and my own development.

For the last three years I have worked with AI daily, and most of that time on my own infrastructure: a GPU server I assembled and tuned myself, with open-source models for text, speech, images and video. I write code together with AI agents. I own the architecture and the decisions; the agents do the routine.

Before AI became my main focus, I spent years on frontend and backend: Angular and NestJS products on MySQL, a SaaS website builder, WordPress sites and integrations for retail and finance. That background is why the AI systems I build ship as working products, not demos.

What I do

AI automation

Agents and pipelines that do a team's routine work from start to finish, with clear inputs, step-by-step checks and a person signing off.

AI infrastructure

Running open-source models on my own GPUs: language models, speech recognition and synthesis, image and video generation, monitoring.

Full-stack development

Web products on Angular and NestJS, MySQL databases, parallel data processing, WordPress, integrations with CRMs and payment systems.

How I work

  1. Business goal first. I start from what the result should change for the business: traffic, conversion, hours saved. Then I choose the tools.
  2. Input decides output. An AI system is only as good as the task setting and data it gets. I design the structure of inputs, the examples and the checks before writing prompts.
  3. Measured, not guessed. Every setting and model choice is backed by a test: speed, quality, cost per run.
  4. Shipped and maintained. Monitoring, safe defaults, documented pitfalls, so the system keeps working after launch.

Selected work

Content and SEO automation for a SaaS company

2025 — present · design and implementation
  • Trained a writing-style model on 425 articles by the company's own authors (LoRA on Qwen 3.5 9B, five iterations in five days at about $4 per run).
  • Articles from the pipeline score 97–100% human on the Pangram AI detector. Measured that choosing the best draft by reading beats choosing by metrics: 3% vs 22% AI.
  • Agent skills cover the whole cycle: topic and brief, writing, keyword research, banners, publishing to the website, versions for five media platforms, performance reports.
  • New landing pages are assembled by an agent from about 55 standard blocks of the live site and arrive as drafts for review.
  • A server (MCP) that starts cloud GPUs only for the length of a batch: no cost while idle, a few cents per batch.
Python, MCP, Claude Code, Hermes Agent, Together AI, WordPress, Playwright

Self-hosted AI lab

2023 — present · built and run by me
  • Assembled a GPU server: 2× RTX 3090 and 2× RTX 3060 (72 GB of video memory), Ryzen 9 7950X, RAID storage, power and fan tuning.
  • Runs a 160-billion-parameter language model with a 128k-token context split across three different GPUs: about 100 tokens per second on average, up to 150.
  • Wrote a gateway that gives every model one OpenAI-compatible address, wakes models on demand, schedules which GPU each one uses and restarts a model that hangs after running out of memory.
  • Speech recognition and synthesis, image, video, music and 3D generation, document OCR, all available to agents through MCP. Existing OpenAI clients work with the local speech service unchanged.
  • Compared inference engines and quantization settings with recorded measurements; maintain my own versioned server builds.
ExLlamaV3, TabbyAPI, vLLM, llama.cpp, Whisper, Docker, Portainer, OpenResty/Lua, Prometheus, Loki, Grafana

Analytics SaaS platform

2022 — 2025 · frontend and full-stack developer
  • User dashboard and the main calculation workflows: results, comparison tables, data export.
  • Server logic and database design for large data volumes.
  • Parallel processing: many machines working with one MySQL/MariaDB database at the same time, with row locking to keep data consistent. Tested storage engines (InnoDB and others) for throughput.
  • A Chrome extension that shows live data from the company's services.
Angular, NestJS, TypeScript, MySQL, MariaDB, RxJS

SaaS website builder

2019 — 2021 · architect and developer
  • A visual builder: websites assembled from ready blocks without programming, still serving client projects.
  • Theme system on CSS variables that reproduces almost any design without code changes.
  • Each site as an isolated project with its own data and publishing; roles and access for different users.
  • Search-friendly single-page sites: pages pre-rendered to HTML for indexing; server-side rendering and a library of 57 web components.
Angular, TypeScript, RxJS, Web Workers, Node.js, Docker

Websites and e-commerce

2019 — 2025 · retail, finance, agencies
  • 25+ websites for small and medium businesses.
  • Online store with automatic price and stock sync from 1C; product bundles.
  • Online booking and payments, CRM integrations, site speed and technical SEO.
WordPress, WooCommerce, PHP, ACF, Tailwind, amoCRM, Stripe

Skills

AI

AI agentsMCPLLM inferenceLoRA fine-tuningQuantizationExLlamaV3vLLMllama.cppWhisper / TTSImage & video generationPython

Frontend

TypeScriptAngularRxJSVueWeb ComponentsTailwind CSSSCSS

Backend and data

NestJSNode.jsPHPREST APIMySQLMariaDBWordPressWooCommerce

Infrastructure

LinuxDockerTraefiknginx / OpenRestyPrometheusGrafanaCloudflareGit

Languages

Russiannative
EnglishB1 · documentation and written communication