AI automation engineer / voice agents / Ukraine, remote

Yevgen Sharyi

I build voice-agent systems that run in production for US companies.

Most recent: a complete voice sales system for a US logistics company. AI agents take inbound and outbound calls, qualify leads and hand them to the sales team; around them I built the telephony, CRM sync, orchestration, the data layer and an admin panel for leads and calls. Before that, eight years in digital marketing, where automating lead handling led me to engineering.

Star map The roadmap.sh AI Engineer roadmap. Lit stars are topics I have used in production work; point at one to see where.

Skill map

The roadmap.sh AI Engineer roadmap, every topic. A filled dot means I have used the topic in real work and can show where; a ring is in progress, a dashed circle is next.

  1. Foundations
  2. Pre-trained models
  3. OpenAI platform
  4. AI safety and ethics
  5. Open source AI
  6. Embeddings and vector databases
  7. RAG
  8. AI agents
  9. Multimodal AI
  10. Development tools
  • shipped
  • in progress
  • next

  1. Foundations

  2. Pre-trained models

  3. OpenAI platform

  4. AI safety and ethics

  5. Open source AI

  6. Embeddings and vector databases

  7. RAG

  8. AI agents

  9. Multimodal AI

  10. Development tools

Selected work

What each system does, what it runs on, and the decisions behind it. Client names are withheld under NDA.

  1. 01US logistics companyOwner, end to endNov 2025 - nowin production

    Voice sales system

    The whole sales loop around the voice agents: inbound and outbound AI calls, CRM sync, orchestration, a Postgres data layer, automatic call review, and an admin panel where the sales team works with leads, calls and agent settings.

    • ElevenLabs Agents
    • Claude
    • n8n
    • Supabase
    • Deno
    • Next.js
    • Twilio
    • Pipedrive
    1. Two generations: a first version with its own test suite and prompt iterations, then the current system from the tender breakdown to live customer calls, with server tools, orchestration, database, admin panels and a separate development environment.
    2. Prices, discounts and order state are decided on the server. The agent can only ask through tools, so nothing a caller says can change a price.
    3. Every tool call is idempotent by its call id, so a retry from the voice platform never creates a second order.
    4. Before each outbound call the system checks the do-not-call list and the lead's local calling window; a lead is claimed atomically in Postgres, so two workers never dial the same person.
    5. A second model reviews every call and may report an error only with a quote from the transcript; critical cases go to the team in Telegram.
  2. 02US business-jet charter operatorTechnical control, fixes, outreachMay 2026 - nowin production

    Pilot sourcing and outreach

    Finds, verifies and contacts business-jet pilots for a charter operator. An external contractor built the first pipeline; I checked it against real vacancies, fixed what was wrong and built the email outreach.

    • Python
    • Streamlit
    • ChromaDB
    • OpenAI
    • Apollo.io
    • FAA registry
    • Mailgun
    1. Checked the returned candidates by hand against the vacancy requirements and traced the irrelevant matches to a stale scraping cache; after the fix the next run returned no irrelevant candidates.
    2. Outreach from the client's own domain with DKIM and DMARC; replies land in the recruiting inbox and get classified; sending stays off until someone switches it on.
    3. Security review of the contractor's API plan: model output passed to eval(), personal data behind unauthenticated endpoints and an open mail relay, each rated critical with a fix.
    4. Contractor work accepted against a written list of acceptance questions and independent testing before the second payment.
  3. 03Voice AI product for US small businessesArchitecture and backendMay - Jul 2026

    AI receptionist from a website

    A visitor enters a website address. The system reads the site, builds a business profile and shows a phone number; the voice agent that answers already knows the business and talks to the caller as its receptionist.

    • ElevenLabs
    • Claude
    • OpenAI
    • n8n
    • Supabase pgvector
    • Twilio
    • Next.js
    1. One persistent voice agent and a small pool of phone numbers, routed by the dialed number when the call starts, instead of a new agent and number per visitor.
    2. Site to profile in two passes: one model reads several pages with web fetch, a second model checks the result; a fallback model keeps the demo reaching a call when the first path fails.
    3. A knowledge base per business in pgvector; abuse limits, consent and a budget cap designed into the first version.
    4. Architecture, the orchestration backend, the admin panel, and a frontend contract for a second developer.
  4. 04Own product, two-person teamLead developerJul - Aug 2026in testing

    Voice interview trainer

    A spoken mock interview against a real job posting for Ukrainian IT candidates, scored, with a session report.

    • Expo
    • React Native
    • TypeScript
    • Gemini Live
    • Vertex AI
    • Supabase
    • Cloud Run
    • RevenueCat
    1. Real-time voice over a raw WebSocket to Gemini Live: voice activity detection tuned for interviews, session resumption, transcription hints for Ukrainian.
    2. Sessions kept dropping under the edge-function runtime limit, so the voice relay moved to a container service with one connection for the whole interview.
    3. Row-level security in every migration, an SSRF guard on the job-URL fetch, and a CV with a prompt injection in the test set.
    4. Expo mobile app with a RevenueCat subscription paywall.
  5. 05Own productSoloAug 2026

    Telegram Mini App with payments

    Tarot readings sold for Telegram Stars: a grammY bot, a Mini App with a 3D deck, and interpretations generated in Ukrainian over a deck drawn by image models.

    • TypeScript
    • Deno
    • grammY
    • React
    • three.js
    • Supabase
    • Vertex AI
    • ComfyUI
    • Vercel
    1. Model chosen by a blind comparison of candidates, with a fallback model behind it.
    2. Payments idempotent on the provider's charge id; user identity taken only from Telegram's own update context, never from message text.
    3. Row-level security and grants written in the same migration as each table.
    4. Deck art from Gemini image models with a FLUX ControlNet fallback; a short-video pipeline on a rented GPU with LTX, Remotion and ffmpeg.
  6. 06Real-estate agency, KyivSole developerApr 2026 - nowin daily use

    Claude Code workspace for a real-estate agency

    Claude Code skills that search property portals, value apartments and rentals, draft listings and build PDF presentations for a small agency.

    • Claude Code
    • Python
    • Playwright
    • SQLite
    1. Portal search through official APIs where they exist and Playwright where they do not; valuation with several statistical methods side by side, plus rental scoring.
    2. A file CRM in JSON and Markdown plus SQLite, readable by the agents and by the agency staff.
    3. Listing forms are filled the way a person would fill them and are never published automatically.
    4. In daily use by the agency since April 2026.

Claude Code

How the systems above get built.

I build with Claude Code: agents write the code, and I own the architecture, the acceptance checks, security, cost and the systems in production.

Process

  1. 01

    Scope

    Which process to automate, and what result will show it worked. Cost is measured before anything scales.

  2. 02

    Spec

    Modules and phases with a definition of done, and acceptance questions written before the build starts.

  3. 03

    Build with agents

    Claude Code writes the code inside the guardrails below; every branch gets a read-only review.

  4. 04

    Run in production

    Test on real calls, then watch quality after launch through automatic call review and alerts.

Practices

  • Context engineering

    Rules per project in CLAUDE.md, persistent memory between sessions, credentials routed by folder with direnv, and MCP servers scoped per project, so an agent only sees its own database.

  • Guardrails with hooks

    PreToolUse hooks that block writes to production hosts and destructive commands and ask before anything ambiguous; a session-start check that warns when privacy settings drift.

  • Skills as procedures

    Repeatable work packaged as skills: property search and valuation for an agency, weekly management reports, n8n workflow builds, telephony and CRM operations that default to read-only.

  • Agent teams and review gates

    Agent teams designed per project, parallel agents for research and implementation, a fact-checker before client deliverables, and a read-only reviewer on every branch.

  • MCP and verification

    Supabase in read-only mode, Playwright for end-to-end checks and screenshots, Context7 for current library docs, Jira and n8n, plus an own MCP server that indexes n8n nodes.

  • Workflow as code

    n8n workflows written in TypeScript with the official SDK, validated over MCP and pushed only to allow-listed workflow ids; database changes only through versioned migrations.

Experience

  1. Nov 2025 - now

    AI engineer, technical lead for AI automation

    US technology holding: logistics software, voice AI products, client projects

    I own the voice-agent direction: I design, build and run the agents, the orchestration and the data layer, and I check contractor deliveries before they are accepted.

    • Voice sales system for the logistics business, end to end, in two generations: the current one went from the tender breakdown to live customer calls, with admin panels and a parallel development environment.
    • Read-only audit of an existing production setup of voice agents and n8n workflows: documented it and ranked the risks, among them unauthenticated webhooks and API tokens stored in the clear.
    • Pilot sourcing for a charter operator: technical control over an external contractor, acceptance before payment, hands-on fixes and the email stack.
    • AI receptionist product: architecture, orchestration backend, vector knowledge base, admin panel, and a frontend contract for a second developer.
    • Aircraft sales and sourcing projects: checked specs and vendor proposals against live FAA registry data; my verdict put a build on hold until the market signals are validated.
    • Delivery estimates for new AI products, benchmarked against published AI-productivity studies; weekly reports to the director in English and Russian.

  2. 2024 - now

    Independent automation developer

    Small businesses, a partner studio, own products

    From 2024, n8n workflows and AI agents for small businesses. Since late 2025, own products built solo or in a pair, each on Supabase with row-level security and migrations from the first commit.

    • n8n workflows with conditional logic, data processing and API calls; integrations between CRMs, Telegram, Instagram, email services and analytics.
    • AI agents and chatbot funnels that qualify leads, segment audiences and answer clients from the first contact to payment.
    • Voice interview trainer and the Telegram Mini App, above.
    • Invoicing app for a partner studio: features, written pull-request reviews, and invoice import from a PDF or photo in a single model call.
    • Browser voice agent with a lip-synced avatar on Gemini Live native audio, server-executed tools and visemes computed from the audio spectrum.
    • Real-estate platform prototype on Twenty CRM, n8n, LangGraph and a custom MCP server, with FreePBX and Fonoster telephony trials.

  3. 2016 - 2024

    Digital marketer

    Two furniture companies, Kyiv

    Marketing end to end: strategy, content, paid ads, landing pages and analytics. Automating lead handling there is what led me to engineering.

    • Paid campaigns on Facebook, Instagram and Google Ads, a landing page per campaign, analytics with GA4, UTM tags and CRM reports per channel.
    • Instagram chatbots that took inquiries, qualified leads and sent reminders, so clients got an answer in minutes instead of hours.
    • Connected the sales team to Instagram inquiries, so every lead went through one pipeline.
    • Short-form video for Reels, TikTok and Shorts: scripts, shooting and editing.

Skills

What AI engineer vacancies ask for most in September 2026, and where I have done it. The full toolset by layer follows.

Asked for in most AI engineer roles

  • LLM APIs and model choice

    Claude, OpenAI and Gemini in production. Models chosen by blind comparison, each with a fallback; reasoning-token cost measured before output limits are set.

    • Claude
    • OpenAI Responses API
    • Gemini
    • Vertex AI

  • Agents and tool calling

    Server-side tools behind voice agents, idempotent by call id, with the server as the last word on prices; a hand-built tool loop over a live WebSocket, no framework.

    • ElevenLabs Agents
    • function calling
    • Gemini Live

  • RAG and vector search

    A knowledge base per business in pgvector, semantic candidate search in ChromaDB, a sales knowledge base in RAG mode. RAG chosen over fine-tuning on purpose.

    • pgvector
    • ChromaDB
    • OpenAI embeddings

  • Evals and quality

    Every production call scored by an LLM judge that must quote the transcript; regression runs with two agents calling each other; test suites by category; blind model comparisons.

    • LLM-as-judge
    • regression runs
    • test suites

  • Prompt engineering and structured output

    Production prompts with mapped call scenarios and tagged sources the agent may quote; strict JSON schemas with per-field rules against hallucination.

    • Structured Outputs
    • JSON Schema

  • Guardrails and AI security

    Caller text treated as untrusted data; injection cases in the test sets; security reviews of AI code, among them model output passed to eval() and personal data behind unauthenticated endpoints; TCPA and DNC rules enforced in code.

    • prompt injection
    • row-level security
    • SSRF guards
    • TCPA and DNC

  • MCP and agentic coding

    Claude Code as the main environment with agent teams, hooks and skills; MCP servers for Supabase, Playwright, Context7, Jira and n8n, and one written in-house.

    • Claude Code
    • MCP

  • Cost and latency

    Cost measured per call minute and per request before scaling; quotas, rate limits and output budgets; the latency of the webhook path measured before deciding where tools run.

    • usage ledger
    • rate limits
    • model fallback

  • Shipping to production

    Live systems with a separate development environment, versioned migrations, feature flags without a redeploy, dry-runs before writes and rollback checkpoints; call review and alerts for monitoring.

    • Docker
    • Cloud Run
    • Vercel
    • GitHub Actions

  • Python and TypeScript

    The systems run on Python (sourcing pipeline, portal scraping, valuation scripts) and TypeScript (edge functions, admin panels, n8n workflows as code). The code is written with Claude Code; I specify, review and test it, with front-end fundamentals from a mentored EPAM program in 2022.

    • Python
    • TypeScript
    • Deno
    • Next.js

Asked for in voice AI and automation roles

  • Voice AI pipeline

    Speech to model to speech with ElevenLabs Agents and Gemini Live: turn-taking, barge-in, voice activity tuning, session resumption; telephony on Twilio with SMS, A2P 10DLC, toll-free verification and answering-machine detection.

    • ElevenLabs
    • Gemini Live
    • Twilio
    • VAD

  • Workflow automation and integrations

    n8n since 2024, now written as code with the SDK and pushed only to allow-listed workflows; webhooks, retries and error paths; CRM, telephony, email, messaging and payment integrations.

    • n8n
    • webhooks
    • Pipedrive
    • Telegram Bot API
    • Mailgun
    • Stripe webhooks

  • Data layer on Postgres

    Supabase Postgres with row-level security from the first migration, least-privilege roles, atomic claims with row locks, state machines in CHECK constraints, scheduled jobs in pg_cron.

    • Postgres
    • Supabase
    • RLS
    • pg_cron

  • Multimodal

    Invoice OCR in a single model call, image generation with a local fallback model, audio pipelines, generated short video on rented GPUs.

    • Gemini image models
    • FLUX
    • LTX video
    • ComfyUI

  • Product and stakeholders

    Business requirements turned into specs with acceptance checks, contractor deliveries accepted against them, weekly reports to management in English and Russian.

Toolset by layer

Models and APIs
  • Claude Opus, Sonnet and Haiku
  • Anthropic web_fetch
  • OpenAI Responses API
  • GPT-4.1 family
  • Structured Outputs
  • text-embedding-3-small
  • Gemini Flash and Pro
  • Gemini Live
  • Gemini TTS and image models
  • Vertex AI
  • Perplexity sonar
Voice
  • ElevenLabs Agents
  • server tools and webhooks
  • language presets
  • Twilio numbers and SMS
  • A2P 10DLC
  • toll-free verification
  • answering-machine detection
  • Gemini Live WebSocket
  • VAD and barge-in
  • whisper
  • Kokoro
  • Simli
Retrieval and data
  • Supabase Postgres
  • RLS and migrations
  • pg_cron and Vault
  • Edge Functions on Deno
  • pgvector
  • ChromaDB
  • ElevenLabs KB in RAG mode
  • SQLite
Orchestration and agents
  • n8n workflow-as-code
  • @n8n/workflow-sdk
  • n8n MCP
  • Claude Code agent teams
  • hooks and skills
  • MCP: Supabase, Playwright, Context7, Jira
  • LangGraph (prototype)
Apps and infrastructure
  • TypeScript
  • Python
  • React
  • Next.js
  • Expo and React Native
  • Vite
  • grammY
  • Telegram Mini Apps
  • three.js
  • Vercel
  • Cloud Run
  • Docker
  • GitHub Actions
  • RunPod
  • Mailgun
  • Pipedrive
  • Playwright
Generative media
  • ComfyUI
  • LTX video
  • FLUX ControlNet via mflux
  • Remotion
  • ffmpeg

Background

Languages
Ukrainian and Russian, native. English is the working language with the US team: specs, reports and vendor calls.
Education
Master's degree, Department of Electronic Systems, Donbas State Technical University, 2008-2013. Front-End Development Program with a mentor, EPAM UpSkill and Strategeast, June to November 2022, diploma ID 913874.
Code
Client code is covered by NDA, so my GitHub repositories (evgensharyy) stay private. Instead of code I offer an architecture walkthrough of the systems above, demos of my own products, or a live build with Claude Code.
LinkedIn
placeholder: profile URL