Evgeny Gmyrya · full-stack developer

I turna few linesinto a workingproduct

Evgeny — an end-to-end web developer from Yekaterinburg. Websites, services, CRMs, Telegram apps and systems of AI agents. In commercial development since 2020.

Nowbuilding agent systems and studying how AI assistants decide what to recommend
(02)Who I am2020 — 2026

For six years I have been building wildly different things. Booking pages for independent masters, with payments. Online stores. Telegram apps where two clients physically cannot take the same slot. CRMs for sales teams. Agent systems that read markets and write copy. Some of it has been running in production for years.

Evgeny Gmyrya, web developer
Evgeny · Yekaterinburgon the web since 2020

An engineering degree, on the web since 2020. Courses gave me the base, real tasks taught the rest. I still read docs and other people’s code — that is the job.

  • 01I own the product end to end — from idea to production
  • 02The stack follows the task, not the other way round
  • 03Agents write code; architecture and review are on me
6+
years of commercial development
54
agent modules in a single system
124
auto-tests in another one
0
external dependencies in a third
(03)What I do05 areas

Five areas I work in

Landings, pages for independent masters, catalogues and stores. A grid and typography of its own for every project, online booking and payments, scroll animation and 3D where they fit.

Design systemsBooking & paymentsSEOWebGL & motion

see → NEW BAND, B2B catalogue

Customer portals, CRMs for sales teams, dashboards and admin panels. Integrations with payment providers, brokers and external APIs, tests and monitoring.

Accounts & rolesDashboardsAPI integrationsTests & monitoring

see → OpenSchool, portfolio terminal

Bots and Mini Apps: booking without sign-up, payments, notifications, broadcasts, Telegram login.

Mini AppsBotsIn-app paymentsNotifications

see → Master booking, “Dispatcher”

Multi-agent systems that hold a conversation, triage requests, read markets and write copy. Local models on my own hardware.

Multi-agent systemsLocal LLMsRAG & memoryAutomation

see → “Nika”, market intelligence

I study how Alice, GigaChat and ChatGPT find and recommend companies and products: markup, feeds, Yandex Commerce Protocol, MCP. I write tools to measure it.

AI visibilitySchema.org & feedsAgentic commerceMCP

see → this site, audit tooling

(04)Selected work06 cases

Six systems. Screenshots, not mockups

Every frame was captured by a script from a running project. Click a screen to open it full size.

01

“Nika” — an AI psychologist

web service · multi-agent system
captured from a live system
6therapy schools in one orchestrator
17screens: chat, radar, tracker
  • An orchestrator classifies the topic and runs the dialogue through six therapy schools (CBT, Jungian, ACT, IFS, narrative, somatic)
  • Crisis detector: on suicide risk it surfaces helpline numbers immediately, no agents involved
  • Subscriptions via YooKassa, Telegram login, PWA, answers streamed over WebSocket
ReactFastAPIWebSocketLLMYooKassaPWA
nika · localhost
“Nika” — an AI psychologist — screen
02

Portfolio terminal

investor dashboard · zero dependencies
captured from a live system
0npm dependencies
48test files
10 000Monte Carlo runs
  • Live data: brokerage account, FRED, SEC EDGAR, Yahoo Finance; a “2 of 3 signals” market verdict
  • Zero npm dependencies: a custom http router, custom math stats (GARCH, HMM, own linear algebra), 48 test files
  • A lab of 13 sections: 10,000-run Monte Carlo, a 4-role LLM committee with Brier scoring, a falsification registry
Node.jsGARCH / HMMMonte-Carlonode:testVPS
terminal · guest mode
Portfolio terminal — before and after redesign — before
before
Portfolio terminal — before and after redesign — after
after
before the redesign → after
03

“Dispatcher” — an AI operator

a personal task operating system
captured from a live system
2interfaces: Telegram and a web panel
1user — me, every day
  • Projects → tasks → “the next physical action”, factoring in energy and time
  • Proactivity with do-not-disturb budgets; long-term memory
  • Fastify 5 + Telegram + GLM + SQLite WAL, Docker, migrations, vitest; ran on my own VPS
FastifyTelegramLLMSQLiteDockervitest
dispatcher · demo
“Dispatcher” — an AI operator — screen
04

NEW BAND — a premium landing

hand-built harmonicas · GSAP
captured from a live system
7parts in the scroll teardown
2languages, leads to Telegram
  • On scroll, the harmonica comes apart into seven labelled parts (GSAP ScrollTrigger + CSS 3D)
  • Two languages; leads go to Telegram via the Bot API
  • Express + PM2 + nginx; responsiveness verified from 320 to 1024+ pixels
GSAPCSS 3DSVGExpressTelegram Bot API
newband · landing
NEW BAND — a premium landing — screen
05

OpenSchool — a school platform

production backend of an online school
captured from a live system
47client views on Vue
prodframes captured straight from production
  • The backend of a working LMS: full-text course search, an admin panel (students, payments, access)
  • Error monitoring, workers, auto-tests
  • Years in operation — screenshots captured straight from production
Node.jsMongoDBElasticsearchAdminBroPM2
openschool · production
OpenSchool — a school platform — screen
06

B2B catalogue of hydrodynamic equipment

a paid commission · catalogue · SEO
captured from a live system
77items collected by a scraper
4sections with a faceted AJAX filter
  • A dealer catalogue without a cart: a faceted AJAX filter with live counters, honeypot-protected leads
  • Full SEO: JSON-LD, OG, sitemap; a Python scraper of the supplier catalogue (77 items)
  • Full documentation handed to the client; still running at the dealer today
WordPressWooCommerceDockerPythonSEO
catalogue · b2b
B2B catalogue of hydrodynamic equipment — screen
(05)Agent engineeringLangGraph · A2A · GigaChat

An agent is a tool, not a wizard

Agents write the code. Architecture, review and responsibility for the result stay with me.

In recent years I have been building systems of AI agents: a multi-agent psychologist with six therapy schools, a 54-module market-intelligence system, a personal operator that lives in Telegram and runs my tasks every day.

That is where the speed comes from. Architecture and review are still done by hand. One level down is my profile too: image segmentation, diffusion inpainting, local LLMs on a home GPU.

Enterprise: LangGraph, the A2A protocol, GigaChat agents with CI/CD — in production at a major bank. Code under NDA.

requestreplycrisisCBTJungACTIFSNarrativeSomaticOrchestrator
01 request02 crisis filter03 classification04 agents05 synthesis
(06)Stack

Tools — to fit the task

Frontend

React/ Next.js/ Vue/ TypeScript/ Three.js/ GSAP/ Tailwind

Backend

Node.js/ Express/ Fastify/ NestJS/ Python/ FastAPI/ PHP/ Bitrix

Data

PostgreSQL/ SQLite/ MongoDB/ Prisma/ Elasticsearch

Agents

LangGraph/ A2A/ GigaChat/ GLM/ Multi-agent/ RAG

Local AI

Local LLM/ SAM/ Stable Diffusion/ GPU inference

Infra

Docker/ nginx/ VPS/ CI/CD/ PM2

Quality

Vitest/ pytest/ node:test/ Code review/ Runbooks

The stack serves the task, never the other way around. Bitrix when the existing infrastructure calls for it. ML pipelines run on my own hardware.

(07)For the curiouslab · index
(08)LinksYekaterinburg ·

Where to find me

Code lives on GitHub. Everything else — Telegram and email.