Senior full-stack engineer:from data model to revenueto productionto monitoringto the user

I own features end to end — data model, API, production UI, infrastructure, monitoring. Production AI is where I go deepest: model cascades, MCP servers wired into internal systems, agentic pre-review.

What I've shipped

1. 01AI assistant for lead processinglead response from an ~8-minute human average to under 30 seconds; after a three-month evaluation comparing human vs. hybrid vs. AI handling, the channel moved fully to the assistant.−
   problem

   managers took about 8 minutes to reply and only during their shift — hot leads went cold, and a chat-manager headcount is expensive.

   constraints

   Load is spiky and the dialogs are live: the operator needs the queue in real time, not on page refresh.

   what I built

   The assistant replies in under 30 seconds at any hour: a lead who writes at night hears back right away, while they still care. I led it: designed the LLM model cascade and routing (cheap model first, expensive on demand) for ~10× lower API costs, Redis caching, and the operator loop — a React admin panel with live dialog monitoring over WebSocket. The operator watches the queue live and steps in before the client notices anything.

   stack

   ReactNodeWebSocket

   Read the write-up (in Russian)Get in touch >
2. 02Push notification servicemarketers run campaigns with zero engineering involvement.+
   problem

   marketing needed to run push campaigns without waiting on developers.

   constraints

   Marketing sends in spikes — thousands of pushes in minutes. Gateways must not fall over, duplicates must not go out.

   what I built

   A template builder and campaign panel in React: a marketer assembles and launches a campaign alone; I scoped it directly with marketing stakeholders. The NestJS backend runs a Redis-backed Bull queue with retries and exponential backoff; sending is idempotent, so duplicates never go out even on retries. Deliverability lives on its own dashboard.

   trade-off

   Direct sending would not survive those bursts, so everything went through the queue. Operations got heavier. The gateways stopped going down.

   what I'd do differently

   Building the editor on contenteditable was a mistake I later cleaned up by hand, edge case by edge case.

   stack

   ReactNestJSRedis Bull

   Read the write-up (in Russian)Get in touch >
3. 03Credit reports in a VK chatbotthe product turned profitable in its first week: net profit +40%, EPC doubled.+
   problem

   open a new sales channel: credit reports issued right in the chat on VK, Russia's biggest social network.

   constraints

   Credit data: one report shown to the wrong person would be one too many.

   what I built

   Led a 3-person team: owned the technical design, user flows and API, and integrated the NBKI credit-history bureau for scoring and decisioning. I designed the data schema for applications, reports and decisions, and built the Node.js backend: search over a 6M-row scoring dataset, the report delivered by phone number right in the bot conversation. Next to the bot, a customer portal where the user sees their applications and decisions.

   trade-off

   The bot and the account area run on one API over one schema. That cost extra design time upfront and keeps the logic in a single place, with no duplication.

   stack

   NodeExpressPostgreSQLVK API

   Read the write-up (in Russian)Get in touch >
4. 04Ad traffic scoring100k-row table: from 8s to under 1s.+
   problem

   9k+ leads a day: impossible to score them and match offers by hand.

   constraints

   Up to 100k rows on a single screen, and the offer decision has to land while the lead is still hot.

   what I built

   A Node + Express microservice polls partner APIs in parallel, with timeouts and retries, so a slow partner never delays the decision. The table is layered: cursor pagination and filtering on the API side, hot queries cached in Redis, react-window keeping ~30 of 100k rows in the DOM, heavy computation in a Web Worker.

   trade-off

   Custom scrolling and selection take more code than dumping every row into the DOM. That's the price of sub-second renders at 60fps.

   what I'd do differently

   Cursor pagination arrived only when the table was already at its ceiling. It belonged there from day one.

   stack

   Reactreact-windowWeb WorkerNodeRedis

   Read the write-up (in Russian)Get in touch >
5. 05Events and notifications on Kafkamessage loss at peaks: zero, and the load is off the main service.+
   problem

   domain events and service notifications were processed inside the main service; at peak load, messages got lost.

   constraints

   what I built

   Moved domain-event processing and service notifications out of the main service into Kafka: producers write events, consumers drain them at their own pace, and the peak spreads out in time instead of hitting the service. Marketing pushes stayed on their Redis Bull queue; Kafka is the event bus. Revenue-critical paths are covered with Playwright end-to-end suites, HTTP with Supertest integration tests, logic with Vitest unit tests; releases go through GitLab CI/CD to Kubernetes, tracing in OpenTelemetry, Grafana dashboards with Telegram alerts — I hear about problems from an alert, not from a complaint.

   trade-off

   Kafka is one more system to deploy and monitor. I took it anyway: a lost message costs more than the extra infrastructure.

   stack

   NodeKafkaPlaywrightSupertestVitestOpenTelemetryGrafana

   Get in touch >
6. 06Catalog Web VitalsLighthouse Performance 58 to 87, conversion rate up about 1.5×.+
   problem

   the catalog was slow: Google ranked it lower, mobile conversion sagged.

   constraints

   Frontend-only changes. The traffic is mobile — every extra kilobyte shows on a weak device.

   what I built

   Split the bundle by route; heavy widgets went into lazy chunks, and preloading stayed only for the critical path. Images got lazy-load and WebP. Load priorities are tuned so the first screen arrives ahead of everything else.

   trade-off

   More requests, and the load priorities have to live in my head. A light first screen on a cheap phone pays for both.

   what I'd do differently

   These days Web Vitals budgets go into CI right away: fixing metrics without a watchdog means fixing them twice.

   stack

   ReactViteLighthouseWebP

   Get in touch >

1. 01 · 2022 — 2023

   SODA

   remote · Aug 2022 — Oct 2023

   Lead-generation agency for residential real-estate developers and law firms: ~60 employees. Owned the client-server side: APIs, database, and interfaces.

   
   role

   Full-Stack Engineer

   product

   AI lead-processing assistant, Node.js microservices, React integration hub, SPA sites receiving paid traffic

   achievements

   • Led the AI lead-processing assistant: lead response under 30 s vs an ~8-minute average; after a three-month humans / humans + AI / AI-only comparison the channel moved fully to the assistant — channel profitability ×3

   • Designed a cascade of LLM models: API costs cut tenfold

   • Rewrote a legacy PHP monolith into Node.js/TypeScript microservices (Express, PostgreSQL): service decomposition, migration of business logic and APIs

   • Internal React integration hub linking the agency's CRM with partner property developers' CRM systems

   • Lighthouse Performance 58 > 87 on the SPA sites receiving the agency's paid ad traffic — conversion up ~1.5×

   ReactTypeScriptNode.jsExpressPostgreSQLLLMLighthouse
2. 02 · 2023 — 2024

   Appbooster

   contract · remote · Nov 2023 — Aug 2024

   Mobile app marketing platform: ~100 employees, 10,000+ apps promoted for clients across 26 countries.

   
   role

   Full-Stack Engineer (contract)

   product

   Task-and-rewards product: React dashboard, chatbot notifications, admin verification queue

   achievements

   • Grew the task-and-rewards product from a few dozen to 1,000 daily active users in 3 months — the whole cycle: a React dashboard for paid tasks and proof, chatbot notifications that bring users back, an admin verification queue

   • Every payout verifiable: the full task lifecycle from pick-up to payout — a moderation queue processing hundreds of completions a day, anti-fraud rules screening out duplicates and faked proofs

   • Ad-traffic attribution (UTM and referrer parsing) alongside activity and return-rate metrics — paid channels judged by retained active users

   ReactNode.jsExpressPostgreSQL
3. 03 · 2024 — now

   Gidfinance

   remote · Sep 2024 — present

   Consumer lending product at fintech company Gidfinance. Cross-functional product team; I ship e2e features from the backend to the interface.

   
   role

   Senior Full-Stack Engineer

   product

   Credit reports, microloan-brand showcases, push campaigns, lead scoring, events on Kafka

   achievements

   • Led a 3-person team on an end-to-end credit-report product — design, flows, API, NBKI bureau, Node.js backend, search over 6M rows; profitable in its first week: net profit +40%, EPC doubled

   • Shipped a feature adjusting microloan-brand showcases to each user's loan history — approvals up

   • Push-campaign builder (NestJS + React): marketers launch automated sends by application status without developer involvement

   • Brought Claude into the project, set up MCP servers, trained colleagues on AI agents — team productivity up

   • 100,000-row lead-scoring table (9K+ leads a day): from 8 s to under 1 s

   • Notification delivery and domain events on Apache Kafka: zero message loss at peak load

   • Devised and built the code-deployment pipeline: GitLab CI/CD into Kubernetes, OpenTelemetry tracing, Grafana alerting

   Node.jsNestJSReactKafkaKubernetesGitLab CI/CDOpenTelemetryClaude CodeMCP