Anuar Beibit
ENRU
Let's talk

Anuar Beibit — AI-native full-stack engineer, Almaty

AI agents write the code. I own what ships.

6 years of product engineering in React, Vue and Node.js. Today Claude Code and my own multi-agent platform do the typing, and nothing reaches production without tests and my review. I set teams up to work the same way: at Mercury, emergency hotfixes dropped 4× in two months.

Discuss a project See the cases
  • 6+ years shipping web products
  • 4× fewer emergency hotfixes after rolling AI into a team
  • 1M+ SEO landing pages generated, in 12 languages
  • 13.6× cheaper agent recon with a local code graph

01 Skills

Skills, drawn as constellations

Five groups that work together on every project, the AI-native core first.

Agents

Teams of agents that plan, code, test and review — each in its own sandbox.

  • AI & Agent Interfaces The layer that turns a model into a system a person can actually operate.
  • Claude Code An agentic coding tool that I work in every day, alongside my own agents.
  • Agent Orchestration The layer that is used to turn one model into a team of cooperating agents.
  • MCP (Model Context Protocol) An open protocol that is used to expose tools, data and prompts to agents in a uniform way.
  • Tool Calling The mechanism that lets a model act instead of only answering.
  • Multi-agent Runs An execution model that splits one task across several agents working in parallel.
  • Sub-agents A pattern that is used to give narrow roles to narrow agents.
  • Sandboxing The isolation layer that decides how much damage an autonomous agent can do.
  • Verification Gates The checks that decide whether an agent’s work is allowed to land.
  • Agent UI & Streaming The interface that makes an autonomous run legible while it is still running.
  • AI Adoption in Teams The work that turns AI tools from a demo into a team's daily practice.
  • Data Anonymization The step that makes AI usable under an NDA.

LLM & evals

Models, context and measurement: quality and cost are measured, not guessed.

  • LLM Application Engineering The craft that is used to build products on top of language models rather than demos.
  • Evals The measurements that show whether a change to prompts, models or the agent setup actually helped.
  • RAG The retrieval layer that gives an agent the right context instead of stuffing the prompt.
  • Code Graph A prebuilt index of how the code is connected that answers structural questions without grep-walking.
  • Anthropic API A model API that drives reasoning and tool use inside the agent loop.
  • OpenAI API A model API that is used as a second provider behind the same agent interface.
  • LiteLLM Routing A multi-provider routing layer that puts one interface in front of many model APIs.
  • Prompt Design The discipline that decides how an agent understands its job.
  • Token & Cost The practice that keeps long autonomous runs economically viable.

Front-end

Interfaces people use every day, from SSR storefronts to real-time agent UIs.

  • TypeScript
  • React
  • Next.js
  • Vue
  • Nuxt
  • Pinia
  • Vite
  • Tailwind CSS
  • Feature-Sliced Design (FSD)
  • SSR
  • JavaScript
  • HTML
  • CSS
  • npm
  • jQuery
  • Angular
  • Svelte
  • Vuex
  • Webpack
  • Jest
  • Jasmine
  • Cypress
  • SPA
  • SSG
  • Three.js
  • WebGL
  • CSS animations
  • Canvas animations
  • Vitest
  • Playwright

Back-end

Services, data and the architecture behind the product.

  • Node.js
  • NestJS
  • Python
  • FastAPI
  • PostgreSQL
  • Redis
  • SQL
  • Ruby
  • Ruby on Rails
  • MySQL
  • Unit Tests
  • E2E Tests
  • Clean Architecture
  • Monolithic Apps
  • Test-Driven Development (TDD)
  • Microservices
  • Domain-Driven Design (DDD)
  • Koa
  • Database denormalization
  • Database normalization
  • Database replication (master-slave)
  • Database indexing

Infrastructure

What keeps it running: containers, CI/CD, servers and networking.

  • Docker
  • Docker Hub
  • Git
  • GitHub
  • GitHub Actions
  • GitLab
  • GitLab CI/CD
  • Webhooks
  • NGINX
  • HTTPS
  • SSH
  • Netplan
  • Proxying

02 How I work

How I work with AI

The pipeline I run every day on my own platform — and set up for teams.

  1. Orchestrator

    A planner splits the task into steps. A stronger model plans and cheaper ones execute — Opus plans, Sonnet works — then the results reconverge.

  2. Agents

    Sub-agents with roles — planner, coder, tester, reviewer — each with its own tools and token budget. Every run gets its own Docker sandbox and git worktree.

  3. Checks & evals

    An agent may push only after typecheck, linters and tests pass. An eval harness runs the same tasks in two setups and compares pass rate, attempts and cost.

  4. Production

    I review the diff and own the release: deploys with a health check and automatic rollback, and no public attack surface.

Rules I keep

  • NDA-safe by default Data is anonymized before it reaches any AI tool.
  • Cost is engineering too A local code graph made agent recon 13.6× cheaper.
  • Proof over promises The first agent-authored PR is merged to main; a real multi-agent task finished with 24 of 24 tests green.

03 Cases

Cases with numbers

Real projects and the numbers they moved.

Mercury Properties · CarCity

AI rolled into an engineering team

4× fewer emergency hotfixes and SSR incidents, in two months

  • Lectures and hands-on sessions first, then the setup that made it stick: editor plugins, custom MCP servers, a RAG index over the code and docs, a local code graph.
  • Data was anonymized before it reached any AI tool — the condition for using AI under the company's NDA.
  • Claude Code
  • MCP
  • RAG
  • Nuxt 3
carcity.kz (opens in a new tab)

Own platform · since 2025

Multi-agent coding platform I use daily

13.6× cheaper agent recon after a local code graph replaced grep-walking

  • Self-hosted fork of OpenHands: planner, coder, tester and reviewer sub-agents, each on its own model, tools and budget.
  • Agents push only through a gate — typecheck, ESLint, Ruff, pytest. The first agent-authored PR is merged to main.
  • React 19 workspace with live agent-event streaming; 2,400+ tests green across 15 locales.
  • React 19
  • TypeScript
  • Python
  • FastAPI
  • Docker
  • LiteLLM
OpenHands, the upstream project (opens in a new tab)

Mercury Properties · MercuryX

Halyk Market integration and a pricing engine

−74% initial load, 2.5 MB → 650 KB

  • Turnkey B2B platform for sourcing and importing from China, part of the Halyk ecosystem.
  • Built the full Halyk Market integration and owned pricing and cost calculation across supplier price, logistics and customs.
  • Rebuilt a barely viable MVP into a modular architecture; shipped sign-in, cart and orders end to end.
  • Nuxt 3
  • TypeScript
  • Pinia
  • SSR
mercuryx.kz (opens in a new tab)

Mercury Properties · CarCity

A marketplace moved to Feature-Sliced Design

+120% feature velocity for the team

  • Migrated an auto-parts marketplace from stock Nuxt to Feature-Sliced Design: less boilerplate, clear module boundaries.
  • Documented architecture, features and business logic — onboarding stopped running on tribal knowledge.
  • Nuxt 3
  • Feature-Sliced Design
  • Vite
CarCity for sellers (opens in a new tab)

POWR.io · remote, US team

Programmatic SEO at scale

1M+ localized landing pages in 12 languages

  • Generated by a pipeline: SEO rankings up, inbound engagement doubled.
  • Customizable website plugins and dynamic UI in React — a 40% lift in plugin sales.
  • Mentored mid and junior engineers across a 12-hour time difference.
  • React
  • Next.js
  • Node.js
  • Ruby on Rails
powr.io (opens in a new tab)

NCRM Group

A CRM front-end with WhatsApp and Telegram inside

2,000+ employees work in the CRM

  • Built the front-end infrastructure and the prototype; brought WhatsApp and Telegram chats into the CRM.
  • Led the sales-conversion statistics module, now used by 50+ companies.
  • Angular
  • Vue.js
  • WebSocket
ncrm.kz (opens in a new tab)

Career

  1. 2025 — now Autonomous Agent Platform AI & Agent Interfaces
  2. 2025 — now Mercury Properties Senior Frontend Engineer
  3. 2024 — 2025 Solution Architects Full-Stack Software Engineer
  4. 2021 — 2024 POWR.io Full-Stack Software Engineer
  5. 2022 — 2023 WoCards.app CTO · Own project
  6. 2020 — 2021 NCRM GROUP Front-end Engineer and Project Manager

04 Formats

Ways to work together

A fixed price for an agreed result, not hours. Start small and scale once it works.

1 week

Test week

One real task from your backlog, done end to end with my agent pipeline. You see the code, the pace and the checks before you commit to more.

  • A task we pick together
  • Code in your repository, with tests
  • A short report: what worked and what's next

Fixed scope

Launch

A project with a clear finish line and a price agreed before work starts.

  • AI in your dev team in about six weeks: agents in the editor, MCP servers for internal systems, RAG over code and docs, NDA-safe data
  • Or an agent system or AI feature, built and brought to a working state
  • Handover with docs, so your team owns the result

Monthly

Support

After launch I stay on, with a monthly scope we agree in advance.

  • New features and code review
  • Agents, prompts and checks kept in shape as models and your code change
  • Costs watched, not guessed
Discuss a project

05 FAQ

Questions, answered

What does “AI-native engineer” mean?

I work with Claude Code and my own agents every day: agents plan, write and test, while I set the task, review every diff and own what ships. Nothing merges until typecheck, linters and tests pass.

Are you an AI engineer or a full-stack developer?

Both, in practice. I build LLM features and AI agents, and I've shipped React, Vue and Node.js products for 6 years. AI is how I work; a working product is what I deliver.

Can you bring AI into our development team under an NDA?

Yes — I did it at Mercury Properties: lectures and hands-on sessions, then editor plugins, custom MCP servers and a RAG index over the code and docs. Data is anonymized before it reaches any AI tool, so client and business information never leaves in identifiable form. Emergency hotfixes dropped 4× in two months.

Which models and tools do you work with?

Claude through the Anthropic API and Claude Code, OpenAI models behind LiteLLM routing, MCP for tools, RAG and a local code graph for context. The model is a runtime choice, not an architectural commitment.

How do you price a project?

A fixed price for an agreed result, settled before work starts. If you'd like to see how I work first, start with a test week.

Do you work remotely? Where are you based?

I'm based in Almaty, Kazakhstan (UTC+5). I spent three years fully remote in a US team at POWR.io across a 12-hour time difference, so remote and async work is routine for me.

06 Contact

Ready to start? Tell me about the task.

Telegram, email or LinkedIn — whichever is easier for you.

Write in Telegram Send an email
Anuar Beibit

Almaty, Kazakhstan · UTC+5