Personal information hidden
This job seeker has chosen to hide his personal information and contact info. You can contact him using this page: https://www.work.ua/resumes/21016267/
AI Engineer (LLM, RAG, MCP)
- City of residence:
- Vinnytsia
- Ready to work:
- Remote
Contact information
Name, contacts and photo are only available to registered employers. To access the candidates' personal information, log in as an employer or sign up.
Uploaded file
This resume is posted as a file. The quick view option may be worse than the original resume.
AI Engineer: LLM Applications, Agentic Systems & RAG
Node.js / TypeScript · MCP · Retrieval & Evaluation · Production Backend
Ukraine, Vinnytsia region · Remote
[
LinkedIn: [
SUMMARY
Backend & AI Engineer with 5+ years of commercial Node.js and TypeScript experience, focused on building LLM-
powered systems that run in production: tool-calling assistants, agentic development workflows and retrieval
(RAG) over source code, including measuring and fixing retrieval quality.
Author of codesearch-mcp, a published MCP server for semantic code search backed by its own retrieval
evaluation harness. At FlexReality shipped an in-product AI assistant to production after a multi-provider
evaluation, and introduced spec-driven Claude Code + MCP workflows that the team used for feature delivery and
bug triage. A strong backend foundation (auth, multi-tenancy, billing, integrations) takes AI features from
prototype to production.
SKILLS
● LLM & Agents: OpenAI API (tool calling), Vercel AI SDK, Claude Code (custom skills, multi-agent review),
MCP server development, MCP integrations (Jira, Sentry, AWS Elasticsearch), prompt design, spec-driven
development (OpenSpec), agentic RAG
● Retrieval & RAG: embeddings, reranking, AST-aware chunking (ts-morph), PostgreSQL + pgvector,
incremental indexing, pluggable local / cloud embedding and reranking backends, HNSW index tuning,
Voyage / Jina embeddings
● Evaluation: retrieval evaluation harnesses (hit@k, MRR@10, per-stage metrics, paired sign tests), labeled
query sets on frozen-commit corpora, model comparison behind feature flags
● Models & Providers: OpenAI, Gemini, Mistral, self-hosted models via Ollama, self-hosted image
generation
● Backend & Data: Node.js (NestJS, Express), TypeScript, REST, GraphQL, WebSockets (Socket.IO), BullMQ,
RabbitMQ; PostgreSQL (pgvector), MongoDB, Redis; TypeORM, Prisma, Drizzle
● Infrastructure & CI/CD: Docker, Kubernetes API, Argo CD, Traefik, GitHub Actions, Bitbucket Pipelines,
AWS (S3, SES), Hetzner, HashiCorp Vault
● Testing & Observability: Jest, Vitest, Sentry, AWS CloudWatch, OpenTelemetry (OpenInference), Arize
Phoenix (LLM / RAG tracing)
● Integrations & Auth: Stripe, LiqPay, App Store / Google Play IAP, LTI, Auth0, OAuth2/JWT
AI PROJECTS
codesearch-mcp · npmjs.com/package/codesearch-mcp · Mar 2026 – Present
● MCP server giving AI coding agents semantic search across local repositories: nine MCP tools (semantic
search, symbol lookup, call and dependency graphs) over PostgreSQL + pgvector, with Nx monorepo
support; ~18k lines of code and ~27.5k lines of tests, built spec-first with OpenSpec.
● Diagnosed empty results on large repositories: pgvector's HNSW scan stopped at 40 candidates
(ef_search) before the project filter and similarity threshold were applied, returning 0 results where 172
qualified. Fixed with per-query ef_search and iterative index scans (pgvector 0.8); tuned index build
parameters (m=32, ef_construction=200), cutting missed nearest neighbors from 1.8% to 0.2%.
● Audited chunking across 76k chunks (67% of class methods had no chunk of their own, 20% had no owning
symbol, ~5.6k fragment chunks) and rebuilt it around declaration boundaries with ts-morph: JSDoc and
decorators stay with their code, plus a linear-time lexer for strings, template literals, JSX and regex.
● Built a reproducible retrieval evaluation framework: hit@k, MRR@10, per-stage metrics (vector vs
reranked), candidate recall and paired sign tests on commit-pinned corpora with index fingerprints.
Changes are gated by acceptance rules set in advance; measurements showed the reranker hurt
description queries (−0.13 MRR@10), which led to a per-call reranker toggle.
● Added a degraded mode that keeps the MCP server running, checks readiness on every call, recovers on
its own and reports the failure cause to the agent. OpenTelemetry tracing with OpenInference attributes,
propagated to child processes and analyzed in Arize Phoenix.
● Used daily on own projects; adopted by part of the FlexReality team and by a few external developers
whose feedback shapes the roadmap.
Spec-driven agentic development toolkit · internal tooling · 2026 – Present
● Extends OpenSpec with custom Claude Code skills for spec writing, spec review and multi-agent code
review; used daily on own projects and previously at FlexReality.
● Agentic RAG at every stage: agents decide when and what to retrieve through codesearch-mcp tools
(semantic search, symbol lookup, call and dependency graphs) while exploring the codebase, writing specs
and reviewing diffs, instead of receiving a fixed context dump.
EXPERIENCE
Backend Engineer · FlexReality · Sep 2024 – Sep 2026
EdTech SaaS platform and internal AI tools
● AI assistant in production: shipped an in-product AI assistant with a frontend engineer (Vercel AI SDK,
OpenAI tool calling): built-in prompts for lesson preparation and formula generation, plus a self-hosted
image generation model integrated as an assistant tool. Evaluated OpenAI, Gemini, Mistral and self-hosted
Ollama models behind a feature flag, demoed to the team and product owners, and rolled out to
production on OpenAI.
● Agentic development workflow: built and ran a spec-driven workflow on OpenSpec and Claude Code with
RAG-backed code exploration, automated spec validation, multi-agent diff review, finding triage and
automated fixes. Delivered the platform auth redesign and platform-wide backend i18n through it.
● Agentic bug triage: built a team-used Claude Code workflow connected to Jira, Sentry and AWS
Elasticsearch via MCP. The agent investigates a selected Sentry error, checks for duplicates and files a Jira
ticket; after human triage, fixes are handed to other agents.
● Introduced the workflow and codesearch-mcp to the team; reviewed backend code and onboarded new
engineers.
● Redesigned authentication and authorization: replaced a monolithic 473-line Express middleware with
~316 declarative, type-safe per-route policies with compile-time coverage checks, and standardized
service-to-service auth across four services in two regions in one coordinated cutover.
● Extracted cross-platform billing into a standalone service and owned the subscription lifecycle across App
Store, Google Play, Stripe and LiqPay: webhooks, idempotency and state reconciliation.
● Built LTI integrations with SSO for Moodle and Ukrainian school systems (NZ.UA, HUMAN); synced users
and billing data with Pipedrive CRM; automated Google Sheets reporting.
Software Engineer · Maven Group · Jan 2024 – Jun 2024
Kubernetes cluster and resource management platform
● Built full-stack (React / Node.js) Backstage.io plugins and REST APIs for Kubernetes cluster management:
programmatic deployments via the Kubernetes API, image builds from GitHub and rollouts through Argo
CD.
Solution Architect / Software Engineer · Rist School · Apr 2023 – Jan 2024
Multi-tenant CRM platform for education
● Sole engineer on a multi-tenant CRM built from scratch (NestJS, Next.js, PostgreSQL, AWS S3 and SES,
HashiCorp Vault): architecture, tenancy model, data isolation, auth and CI/CD on Bitbucket, from business
analysis to deployment.
Node.js Developer · QATestLab, BezBrendu · Jan 2021 – Feb 2023
● Corporate chatbot on Microsoft Bot Framework, integrated with internal systems (QATestLab);
Facebook / Instagram Marketing API SDK, REST APIs and Auth0 permission sync for a sales CRM
(BezBrendu).
EDUCATION & TRAINING
● Vinnytsia Technical College, Radio Engineering · Certificate: AI agent context workshop
Languages: English – Intermediate (B1) · Ukrainian – native
More resumes of this candidate
Remote
OLEKSANDR ANTONIUK Backend Engineer Node.js / TypeScript · Backend Architecture · Agentic AI & RAG Ukraine, Vinnytsia region · Remote · LinkedIn: · Telegram: @underground_alex SUMMARY Backend...
Similar candidates
-
AI-креатор, AI Artist
Remote -
AI engineer
Remote, Other countries -
LLM Engineer
Remote, Ivano-Frankivsk -
AI Automation Engineer
Remote -
AI Prompt Engineer
62000 UAH, Remote, Kyiv -
Спеціаліст з ШІ-автоматизації бізнесу, Prompt Engineer (AI Architect)
Remote, Dnipro , more 5 cities