Nikita

Nikita Linchak
[відкрити контакти](див. вище в блоці «контактна інформація») | Telegram: @nikos3rd | [відкрити контакти](див. вище в блоці «контактна інформація») | github.com/nikos3rd
Kyiv, Ukraine | Open to remote / relocation

SUMMARY
I build LLM systems that survive contact with production. In practice that means retrieval that returns the right clause rather than a similar
paragraph, agents that cannot take an action their schema does not permit, and pipelines where a repeated run never does completed
work twice. Most recently I worked across three production systems - an internal HR platform in strict TypeScript on AWS, an AI outreach
service over a third-party platform with no public API, and a LangGraph lead engine. Before that I shipped 96 production automation
workflows across 13 business areas - lead scoring, CRM, content and sales operations included - and an asynchronous RAG platform
over four vector stores, while running the Linux and Windows infrastructure underneath. Usually the only engineer on that surface, from
process analysis through deployment, monitoring and the troubleshooting that follows.
PROFESSIONAL EXPERIENCE

AI Product Engineer | Growth Automation & Product Operations
June 2026 - September 2026 | Product company - internal platforms and AI outreach systems
Worked across three production systems: an internal HR platform, an AI outreach service built over a third-party platform with no public
API, and a LangGraph lead-generation engine. Client details are under NDA; the descriptions below cover architecture and engineering
decisions only.
HR platform - TypeScript, Express, MongoDB, React 19, AWS
• Strict-TypeScript modular monolith on Express and Mongoose serving a same-origin React 19 SPA, exposing 200+ generated RPC
methods and covered by 400+ test files across unit, integration and cross-browser Playwright end-to-end suites.
• Built the domain: a multi-type absence ledger with decimal-safe accrual and approval hierarchies, a generic requests hub with optimistic
locking, a drag-and-drop workflow designer driven by employee lifecycle triggers, a reporting suite exporting CSV and XLSX, RBAC with
field-level security, and full bilingual localisation across 30+ screens.
• Made it safe to operate on sensitive personnel data: tamper-evident hash-chain audit logging with external anchoring, append-only
ledgers, a transactional outbox for guaranteed notification delivery, corporate SSO restricted by domain allowlist with no public signup,
and schema validation that refuses to boot on a missing variable.
• Owned deployment on AWS: containerised multi-service stack from a commit-pinned immutable image, ordered bootstrap of database,
migrations, seeding and application services, rollback by image tag, an operational runbook and a production readiness audit.
AI outreach service - Python, FastAPI, PostgreSQL
• Integrated a third-party outreach platform that publishes no API, by reverse-engineering its network traffic into a config-driven endpoint
map, with session lifecycle handling (proactive and reactive refresh, automated re-login), 401 retry, backoff and request pacing.
• Enforced single-session egress with a PostgreSQL advisory lock, since the platform permits one session per account, and deliberately
ran the webhook receiver, approval console, queue worker and scheduler in one process so that lock could not be bypassed.
• Put a person in front of every irreversible step: a cost and volume forecast before launch, a human approval flow (console plus
Telegram), guardrails for working hours, sending pace, price talk and AI disclosure under Article 50 of the EU AI Act, and mandatory
human routing whenever a lead showed real interest, raised price, reacted negatively or asked whether it was a bot.
• Kept the MCP server out of the production image on purpose: it exposes unauthenticated tools capable of messaging real people, so it
runs bound to loopback on a developer machine instead.
Lead engine - Python, LangGraph, FastAPI, PostgreSQL
• LangGraph pipeline from intake through enrichment, ICP qualification, a review gate, service-fit scoring and a structured plan to
grounded drafts, approval and export, checkpointed with PostgresSaver so an interrupted run resumes instead of restarting.
• Grounded generation contract: drafts may use verified evidence only, review workbooks are marked REVIEW ONLY, and a sequence
becomes sendable only for the exact approved revision.
• Shared prospect registry that never re-purchases an enrichment record the team already owns, merges canonical identities on
duplicate email and keeps durable reservations, cutting third-party provider spend directly.
• Fail-closed by design: an unavailable provider returns an explicit error rather than degrading quietly, and production migrations run
behind an automated backup-then-migrate chain.
Stack: TypeScript, Node.js, Express, Mongoose, MongoDB, React 19, Vite, Zod, Vitest, Playwright, Python, FastAPI, PostgreSQL,
LangGraph, Alembic, Claude API, Docker Compose, GitHub Actions
AI Automation Engineer and System Administrator
January 2024 - July 2026 | Sole AI specialist - automation, HRTech, sales, research and infrastructure
Joined as a System Administrator, then grew into AI and business process automation in the same company while keeping the IT
infrastructure running. No platform team and no QA: whatever I shipped, I also operated.
AI & automation
• RAG platform. Asynchronous indexing with automatic re-embedding when a source document changed, so the index never drifted
silently from the documents. Hybrid retrieval combining BM25 with Sentence-BERT embeddings and HNSW, merged through
Reciprocal Rank Fusion and refined by a cross-encoder over the top candidates only - adopted after pure vector search proved
unreliable on exact terms and identifiers. Multi-vector-store architecture across Qdrant, Pinecone, pgvector and Supabase, orchestrated
with LangChain.
• 96 production workflows in n8n across 13 business areas. Built with retries, idempotent reruns, error handling and operational
alerting, because a silently broken automation costs more than the manual work it replaced.
• Production AI agents. Semantic email classification and routing, automated CV screening with candidate-to-vacancy matching and
structured reasoning for recruiters, and a competitive intelligence agent over large volumes of research data. On top of them, a
multi-agent Telegram assistant: central orchestration over dedicated research, email, document and memory agents, with
Pinecone-backed long-term context shared between them.

• Custom MCP servers. Servers exposing GitHub and Docker operations to Claude and to n8n workflows, including one built over a
private API reverse-engineered from its network traffic, with session management, rate limiting, access controls and documented
behaviour.
• AI lead scoring. Lead CRM backend on FastAPI and PostgreSQL with asynchronous processing and an LLM scoring engine that
persisted its rationale, so a salesperson could audit why a lead was ranked as it was instead of trusting a number.
• Content automation. End-to-end pipeline across several generative platforms: idea generation, creation, quality review, approval,
scheduling and publishing, with monitoring around external providers.
Systems administration & IT operations
• Windows, Linux & identity. Administered Windows Server clusters and Debian/Ubuntu servers; Active Directory, Group Policy, DNS
and DHCP. Configured NGINX and Apache for corporate web services and ran Docker-based production services.
• Virtualisation & backup. Maintained VMware ESXi and Proxmox environments with backup and disaster recovery on Veeam Backup
& Replication.
• Networks & communications. VLAN segmentation on Cisco and MikroTik, L2TP/IPsec remote access for distributed employees,
structured cabling and patch panels, Synology NAS and Grandstream VoIP.
• Web & security testing. Led the corporate website from Figma designs and custom-coded prototypes to WordPress production; tested
APIs with Postman and assessed web vulnerabilities with Burp Suite.
• Support. Tier 2/3 support for 50+ users, including workstation hardware, SolidWorks and 1C ERP.
Stack: Python, FastAPI, PostgreSQL, Node.js, n8n, LangChain, RAG, MCP, Qdrant, Pinecone, pgvector, Supabase, OpenAI API, Claude
API, Gemini, Telegram API, Slack API, Docker, Linux, Windows Server, VMware ESXi, Proxmox, NGINX, GitHub Actions
AI Automation Engineer (contract, part-time)
January 2026 - May 2026 | International e-commerce - marketplace automation and data operations
• 76-node listing pipeline. Designed and built the production n8n workflow that rewrites and publishes product titles at scale across
several Latin American markets, using Gemini 2.5 Flash through Vertex AI with BigQuery staging tables as the source of truth.
Consolidated fragmented legacy automations into one pipeline.
• Safe reruns and cost control. Idempotent execution on top of BigQuery so a repeated or resumed run never reprocessed completed
work, plus caching that prevented duplicate LLM calls for repeated products.
• Marketplace integrations. Mercado Libre and Amazon Selling Partner APIs with OAuth 2.0, multi-market Amazon Buy Box monitoring,
and scheduled Slack reports on status and failures.
• Analysed workflow performance and external API bottlenecks, optimised batch processing, and documented the architecture and
recommendations for the operations team.
Stack: n8n, Gemini 2.5 Flash, Google Vertex AI, BigQuery, Mercado Libre API, Amazon Selling Partner API, Slack API, OAuth 2.0, Python

PROJECTS
Hybrid Information Retrieval System - personal R&D
• BM25 lexical retrieval plus Sentence-BERT embeddings and HNSW, merged by Reciprocal Rank Fusion and reranked by a
cross-encoder over the top candidates, with a Next.js application for result inspection and side-by-side relevance comparison.
Self-hosted Kubernetes Platform - personal infrastructure lab
• GitOps platform I run myself on Proxmox VE: Fedora CoreOS nodes provisioned with OpenTofu and Ansible, k3s migrating to a highly
available kubeadm cluster, Argo CD with ksops, SOPS and age for secrets, kube-prometheus-stack and Loki, cert-manager for TLS,
Forgejo and MinIO for source and object storage, Tailscale for access. Built to learn production Kubernetes and GitOps properly rather
than from a tutorial.
TECHNICAL SKILLS
AI & agents: Claude API, OpenAI API, Gemini, Vertex AI, LangChain, LangGraph, Model Context Protocol, RAG, multi-agent
orchestration, structured outputs with deterministic validation, prompt engineering, LLM evaluation, Claude Code, Cursor
Automation & workflows: n8n, Make, Zapier, Airtable, webhooks, Slack, Telegram, Google Workspace, Stripe API, Twilio, Perplexity
Email & LinkedIn outreach: Closely, lemlist, Instantly, Salesforge, La Growth Machine, Linked Helper, Mailchimp
Lead generation & enrichment: Apollo, Clay, LinkedIn Sales Navigator, Snov.io, Hunter, PhantomBuster, Apify, Firecrawl
CRM & sales operations: HubSpot, Pipedrive, Salesforce, AI lead scoring, pipeline and CRM automation
Content, social & SEO: Blotato, Manychat, Canva, ElevenLabs, Google Search Console, Semrush
Project & product operations: Technical project management, product marketing support, ClickUp, Notion, Jira, Trello
Backend & frontend: Python, FastAPI, TypeScript, Node.js, Express, React, Next.js, Tailwind, PostgreSQL, MongoDB, SQL, Alembic,
REST and RPC APIs, OAuth 2.0, async processing, Zod, Pytest, Vitest, Playwright
Retrieval & data: Qdrant, Pinecone, pgvector, Supabase, BM25, Sentence-BERT, HNSW, Reciprocal Rank Fusion, cross-encoder
re-ranking, re-embedding pipelines, BigQuery, Mercado Libre API, Amazon Selling Partner API
Cloud & DevOps: AWS, Google Cloud, Docker and Docker Compose, GitHub Actions, Kubernetes, Argo CD, OpenTofu / Terraform,
Ansible, Prometheus, Loki, SOPS, Linux, NGINX, Apache
Systems & networking: Windows Server, Active Directory, GPO, DNS, DHCP, VMware ESXi, Proxmox, Veeam, Synology, Cisco,
MikroTik, VLANs, L2TP/IPsec, Grandstream VoIP, Postman, Burp Suite

EDUCATION, CERTIFICATIONS AND LANGUAGES
BSc in Computer Science - Kyiv National University of Technologies and Design, 2022 - 2026
Certifications: Creation and Development of IT Products - Genesis Tech Academy; Web Technologies Student Competition Diploma -
KNUTD
Languages: Ukrainian - Native; Russian - Native; English - Upper-Intermediate (B2); German - Beginner