AI Engineer / Autonomous Agent Systems / Remote

Yevgen Sharyi

AI software engineer designing agentic systems that a business trusts with money and customers.

AI engineer with live customer-facing systems in production since 2025. I bridge the gap between volatile LLM outputs and deterministic business execution: architecting server-authoritative tool loops, telephony pipelines, strict Postgres state machines, and automated transcript audits. The backend platforms and front ends around them are mine too: admin portals, dashboards and mobile apps on Next.js, React Native and Postgres. Currently running multi-tenant voice agents and an enterprise marketing platform across logistics and aviation. Backed by eight years in commercial marketing, every system is designed around conversion, unit economics, and zero financial risk.

NowEngineering and operating live production agents across three businesses and a multi-tenant platform. Developing with agentic AI developer tools under strict automated guardrails. Available for remote AI engineering roles across the US and Europe.

Selected Work

A mix of production work and personal builds: AI agents, voice systems, backend platforms and mobile apps. Some are live, some are in development, a few are internal. Clients are covered under NDA, so each deployment is described by domain. Every architectural claim below is verifiable in an interview.

  1. 01US logistics enterprise, then two private aviation platformsLead Systems Architect & DeveloperNov 2025 - nowin production

    Voice Sales System

    Sub-minute lead callback, inbound call handling, qualification and order creation directly in core business systems. ElevenLabs, Twilio, Edge Functions and Postgres, with an operator dashboard and an independent LLM judge scoring every transcript. Built for a US logistics enterprise, then deployed on the same architecture for a private jet charter operator and an aviation software platform.

    1. Second-generation platform from scratch: compiled Edge Functions instead of experimental workflows, centralized Postgres state, a staging environment matching production.
    2. Replicated for the private jet operator in 8 days; ported to a membership aviation platform in 1 day via RAG price lookup with strict parity verification.
    3. Server-authoritative governance: prices, discounts and order states live only in database constraints; the agent reads them via read-only tools, so prompt manipulation of financial terms is impossible.
    4. Idempotent execution: every tool invocation is keyed by voice session ID, so carrier webhook retries never duplicate an order.
    5. Telephony safeguards: TCPA compliance, local time-zone calling windows, do-not-call registries and atomic row-level locks in Postgres against dual-agent dialing.
    6. Independent quality judge: a secondary model audits every transcript and needs verbatim citations to flag non-compliance or unstated price caveats. Critical flags alert on-call teams in Telegram.
    7. Trilingual agent: English, Ukrainian and Russian, with deterministic rollback checkpoints before each release.
    • ElevenLabs Agents
    • Claude
    • Supabase
    • Deno Edge Functions
    • Postgres
    • Next.js
    • Twilio
    • Pipedrive
  2. 02Charter platform for operators, brokers and aircraft owners, USPilot module: backend and frontendMay 2026 - nowin production

    Pilot Sourcing and Outreach

    Multi-tenant charter platform for operators, brokers and aircraft owners; I built its pilot module end to end, backend and frontend. A vacancy is parsed into criteria, candidates come from registries, job boards and enrichment services, every one is verified against the FAA registry and ranked for the role, then an outreach sequence goes out from the company's own domain. Replies and job board applications are collected automatically and read by AI: interest, qualification, next step.

    1. Pilot pool shared by the whole company: FAA verdict with its reason, hours, type ratings, contacts, do-not-contact flags and per-vacancy verdicts on one card.
    2. Sourcing: vacancy criteria, search runs, FAA verification, contact enrichment through Apollo.io and LinkedIn lookup, ranking against the role.
    3. Outreach: campaigns from a dedicated sending domain with DKIM, SPF and DMARC; replies routed back into the system and classified by a model into interest and qualification.
    4. Applications: the vacancy is published on a pilot job board through a browser container, applicants are collected and verified, job board credentials are encrypted at rest.
    5. Security review before release: removed eval() on model output, an unauthenticated PII endpoint and an open mail relay; acceptance suites and a dry run before each module goes live.
    • Python
    • FastAPI
    • PostgreSQL
    • Celery
    • OpenAI
    • Playwright
    • Next.js
    • Apollo.io
    • FAA registry
    • Mailgun
  3. 03US technology holding portfolioPlatform Architect & Backend LeadSep 2026 - nowin production

    Marketing Platform

    Multi-tenant engine where AI agents draft strategies, articles, app store metadata, ad copy and customer responses for an enterprise portfolio; operators review in an admin portal, and server-side transactional gates keep unapproved drafts from ever executing externally. It also powers a done-for-you AI marketing agency: agents and marketers handle SEO, Google Ads, social media, email and reviews 24/7, and the client sees every post, lead and result live in their own dashboard. In production since October 2026.

    1. Race-condition prevention: publishing runs under database row locks, with third-party API dispatch decoupled from the transaction, so nothing double-posts.
    2. Transactional budget controls: spend limits validated in the exact transaction that records approval, backed by concurrency stress tests that try to double-spend.
    3. Strict tenant isolation: every query enforces ownership server-side; cross-tenant probes get a 404.
    4. Cryptographic key isolation: ad and social tokens encrypted with keys stored outside the database; interfaces write tokens and can never read them back.
    5. Untrusted content sanitization: reviews, competitor pages and LLM completions are untrusted data; models never select payees or financial actions, and every output passes schema validation.
    6. Deployment discipline: connectors in mock mode locally and in CI/CD; staged dev-to-prod releases governed by custom skills of the agentic AI developer tool.
    • Python
    • FastAPI
    • SQLAlchemy
    • PostgreSQL
    • Celery
    • Claude
    • Next.js
    • shadcn/ui
    • GitLab CI
  4. 04Voice AI product for US small businessesArchitecture & Backend LeadMay 2026 - nowin production

    AI Receptionist from URL

    Turns any business URL into an operational inbound voice receptionist in seconds: the pipeline crawls the website in tiers, extracts operational boundaries into pgvector, provisions dynamic Twilio routing and answers callers without a separate agent or number per user.

    1. Scalable telephony pool: dynamic inbound routing over a shared phone pool, matching caller profiles at session start without dedicated numbers per demo.
    2. Dual-pass extraction: async crawlers aggregate business facts while a validation model filters hallucinations; fallback models rule out dead-ends.
    3. Tenant isolation: individual vector partitions in pgvector, abuse rate-limits, explicit caller consent and hard budget caps.
    4. Modular contracts: orchestration backend, admin controls and formal API schemas for frontend integration.
    • ElevenLabs
    • Claude
    • OpenAI
    • Supabase pgvector
    • Twilio
    • Next.js
  5. 05Proprietary mobile product, two-engineer teamLead Developer & Systems EngineerJul - Aug 2026in development

    Real-Time Voice Interview Simulator

    Spoken technical interview simulator tailored to live job vacancies for European and Ukrainian IT talent: sub-second conversational latency, structured competency scoring and full post-interview feedback.

    1. Raw WebSocket voice pipeline to Gemini Live with tuned VAD, session resumption and Ukrainian phonetic hints.
    2. Persistent container relay on Cloud Run holds WebSocket connections open for 45-minute sessions, removing edge runtime timeout drops.
    3. Defensive posture: Row-Level Security in every migration, SSRF filtering on vacancy ingestion, adversarial prompt injection tests.
    4. Cross-platform Expo React Native app with auto-renewing subscriptions via RevenueCat.
    • Expo
    • React Native
    • TypeScript
    • Gemini Live
    • Vertex AI
    • Supabase
    • Cloud Run
    • RevenueCat
  6. 06Proprietary mobile productProduct Owner & DeveloperApr 2026 - nowin production

    Mobile Invoicing App for Freelancers

    Mobile-first invoicing for freelancers: proposal, contract and invoice in one flow, sent from the phone, with open tracking and overdue follow-ups. An existing invoice is imported from a PDF or a photo in one model call. Two visual languages on purpose: the app for the freelancer, a clean PDF for the client.

    1. Proposal, contract and invoice share one data model, so a change in the proposal flows into the invoice.
    2. Import from a photo or PDF is one structured-output call, validated with zod before anything is saved.
    3. Supabase auth, database, storage and edge functions with Row-Level Security; subscriptions through RevenueCat.
    4. PDF rendered on the device with expo-print, charts with victory-native, motion in Reanimated and Skia.
    • React Native
    • Expo
    • TypeScript
    • Zustand
    • React Hook Form
    • zod
    • Supabase
    • RevenueCat
    • Reanimated
    • Skia
  7. 07Proprietary consumer productSole DeveloperAug 2026in testing

    Telegram Mini App with In-App Economy

    Interactive 3D application with payments in Telegram Stars, automated multi-card interpretations and generative visuals produced on rented GPU clusters.

    1. Objective model benchmarking: production LLMs chosen by blind evaluation on test sets, with automatic secondary fallbacks.
    2. Financial idempotency: fulfillment locked to provider charge IDs; identity resolved only through Telegram cryptographic update signatures.
    3. Zero-trust database: RLS policies and grants declared in the same migration as each table.
    4. Generative media: card artwork via Gemini image pipelines with local FLUX ControlNet failover; short videos via Remotion and headless ffmpeg on rented GPUs.
    • TypeScript
    • Deno
    • grammY
    • React
    • three.js
    • Supabase
    • Vertex AI
    • ComfyUI
    • Vercel
  8. 08Real-estate brokerage, KyivSystems Architect & DeveloperApr 2026 - nowin daily use

    Autonomous Brokerage Workspace

    An agentic AI developer tool plugged into a boutique brokerage's daily work: a CRM of clients, properties and deals that the agent reads and updates, portal search and daily monitoring, comparative sale and rental valuations, contracts, owner proposals and client presentations generated from the data. In commercial use every business day since April 2026.

    1. CRM for the agent: JSON tables plus Markdown cards with a relation schema; every change is validated against it (links exist, ids unique, numbers recomputed).
    2. Hybrid ingestion: official portal APIs plus Playwright automation for unindexed listings; valuations compare several pricing approaches.
    3. Generators: contracts, owner proposals, property and investor presentations with payback, as HTML and PDF.
    4. Human approval gate: automated drafts, mandatory human verification before external syndication.
    • agentic AI developer tool
    • Python
    • Playwright
    • SQLite

Engineering Methodology

Disciplined harness for agentic coding and autonomous systems.

Agentic AI developer tools implement the codebase inside boundaries I construct. I design the system architecture, write the formal specifications, enforce automated security hooks, and own production reliability. The speed is multiplied; the accountability is 100% mine.

Process

  1. 01

    Scope & Unit Economics

    Target the high-friction workflow. Measure token budgets, latency budgets, and conversion metrics before drafting code.

  2. 02

    Formal Specification

    Define interface contracts, state machines, and acceptance test matrices before invoking coding agents.

  3. 03

    Guarded Implementation

    The agentic AI developer tool executes implementation inside pre-tool hook boundaries. Every branch undergoes automated read-only review.

  4. 04

    Production Hardening

    Validate with real call traffic. Continuous transcript monitoring via secondary judge models and real-time alert loops.

Principles

  • AI where it earns its place

    A deterministic rule beats a model call whenever the rule is enough. The model gets the parts where language is the actual problem.

  • Design the system before the screen

    Tables, states and failure modes first. The UI comes last and inherits the constraints.

  • The server has the last word

    Prices, limits and order states are database constraints. The agent reads them and never sets them, whatever the caller says.

  • Own the prompt and the tool surface

    Structured outputs over free text, schema-grounded prompts over guesswork, confirm-before-invoke on anything destructive. No chain of abstractions I cannot debug.

  • Build for the real world

    Retries, idempotency keys, timeouts, dead-letter queues and staging parity. Webhooks arrive twice and connections drop, and the system has to carry on.

  • Stay close to the craft

    I read the code the agent writes, review every diff and keep repeatable procedures as versioned skills, so the next project starts from what worked.

Experience

  1. Nov 2025 - now

    AI Engineer, Technical Lead for AI Systems & Automation

    US technology holding: logistics software, voice AI products, enterprise client solutions

    Own the AI systems of a US technology holding and its enterprise clients: directing end-to-end architecture, backend infrastructure, live operations, and safety audits across all customer-facing AI products.

    • Architected and shipped the production voice sales system for a US logistics enterprise, handling live inbound and callback customer calls.
    • Scaled the identical voice architecture to two aviation platforms (private jet charter operator in 8 days, aviation SaaS in 1 day) with 100% functional parity.
    • Designed and shipped the enterprise marketing platform from scratch: multi-agent generation, human approval portals, and transactional publication gates. Live since October 2026.
    • Conducted comprehensive security audits of legacy production voice setups, identifying and remediating unauthenticated webhooks and plaintext credential storage.
    • Engineered the pilot sourcing and regulatory outreach pipeline: integrated FAA pilot database queries, acceptance suites, and authenticated email deliverability.
    • Audited vendor proposals for aircraft sourcing against live FAA registry data; technical evaluation prevented an unviable multi-month build.
    • Architected the AI receptionist SaaS: multi-tier web ingestion pipeline, vector knowledge bases in pgvector, and multi-tenant admin dashboards.
    • Directed delivery estimations and presented weekly technical and operational progress to executive leadership.

  2. 2024 - 2025

    Independent AI Systems Developer

    Small businesses, partner engineering studios, proprietary SaaS products

    Architecting production automation and standalone AI products. Full-stack TypeScript/Python applications on Supabase with Row-Level Security enforced from initial migration.

    • Engineered resilient event-driven integrations linking CRM systems, Telegram, email, and analytics pipelines.
    • Designed autonomous qualification bots managing end-to-end customer onboarding from inquiry to payment.
    • Delivered invoicing system for partner design studio: code reviews, automated invoice OCR extraction in a single model call.
    • Engineered real-time browser voice agent with animated conversational avatar on Gemini Live with server-side tool execution.
    • Prototyped enterprise real-estate automation engine integrating Twenty CRM, LangGraph, and custom MCP servers.

  3. 2016 - 2024

    Head of Digital Marketing / Growth Lead

    High-ticket manufacturing and retail brands, Kyiv

    Led digital customer acquisition and marketing infrastructure: paid media, funnel architecture, automated lead pipelines, and web analytics. This eight-year foundation in commercial conversion is why every AI system I build today is grounded in unit economics, customer trust, and business ROI.

    • Managed multi-channel performance advertising across Google Ads, Meta, and native networks with end-to-end CRM attribution.
    • Built conversational qualification funnels in social channels, automating lead routing from initial touchpoint to sales closers.
    • Unified fragmented lead channels into centralized operational pipelines, cutting lead response times.
    • Directed high-volume creative production and short-form video pipelines from narrative strategy to editing.

Skills

Verified engineering stack and production operational track record.

Agentic AI

  • LLM APIs & Architecture

    Claude, OpenAI Responses API, and Gemini in live production. Model selection governed by blind benchmark evaluations, automated fallback orchestration, and token budget enforcement.

    • Claude Opus/Sonnet
    • OpenAI Responses API
    • Gemini Live
    • Vertex AI

  • Agents & Tool Execution

    Server-authoritative tool calling behind voice and text agents, strictly idempotent by invocation ID. Custom WebSocket tool execution loops operating without framework bloat.

    • ElevenLabs Agents
    • Function Calling
    • Gemini Live WebSockets

  • Multi-agent systems

    Dual-model supervision: independent judge models, fact-checking agents, and isolated code review reviewers.

    • LLM-as-judge
    • fact-checker agents
    • read-only reviewers

  • Prompt Engineering & Schemas

    Production system instructions with deterministic state transitions and source attribution. JSON Schema enforcement with field-level regex rules preventing hallucinations.

    • Structured Outputs
    • JSON Schema
    • State Mapping

  • RAG & Vector Databases

    Tenant-isolated vector stores in pgvector, semantic candidate search in Postgres, and real-time knowledge base retrieval for pricing lookups. Focus on clean data pipelines over complex fine-tuning.

    • pgvector
    • OpenAI Embeddings

  • Evaluations & Quality Control

    Automated LLM-as-judge auditing every production call with mandatory transcript citations. Regression test suites, multi-agent adversarial testing, and blind comparison runs.

    • LLM-as-Judge
    • Regression Testing
    • Adversarial Evaluation

  • Security & Defensive AI

    Zero-trust input sanitization for caller speech, web scrapes, and model completions. Security reviews intercepting prompt injection, remote eval execution, and unauthenticated endpoints. Code-level TCPA and DNC enforcement.

    • Prompt Injection Defense
    • Row-Level Security
    • SSRF Protection
    • TCPA Compliance

  • AI governance

    Risk tiers per agent profile; mandatory human authorization gates on payments, messaging, and database updates.

    • risk tiers
    • human approval gates

  • Observability

    PII-sanitized telemetry, append-only operational audit trails, latency percentiles, and error dispatch dashboards.

    • PII-sanitized telemetry
    • audit trails
    • latency percentiles

  • Cost & Latency Engineering

    Real-time ledger accounting for call minutes and token consumption. Hard quotas, rate limiters, model fallbacks, and webhook latency benchmarking prior to architecture decisions.

    • Token Ledgers
    • Rate Limiting
    • Sub-second Latency

  • Agentic Coding & MCP

    An agentic AI developer tool as the primary development environment: custom agent teams, pre-tool hooks, and automated skills. MCP servers for Supabase, Playwright, Context7, Jira, and custom engines.

    • agentic AI developer tool
    • Model Context Protocol (MCP)
    • Hooks

Voice, data and delivery

  • Voice AI Pipeline

    Low-latency speech-to-speech pipelines using ElevenLabs Agents and Gemini Live: millisecond turn-taking, barge-in tuning, VAD calibration, and session resumption. Telephony via Twilio with A2P 10DLC registration, toll-free verification, and answering-machine detection.

    • ElevenLabs
    • Gemini Live
    • Twilio
    • VAD & Barge-in

  • Workflow Automation & Integrations

    SDK-based automation as code, deployed strictly to allow-listed workflows. Resilient webhook listeners, exponential retries, and dead-letter error handling across CRMs, email, and billing.

    • n8n as code
    • Pipedrive
    • Telegram Bot API
    • Mailgun
    • Stripe

  • Data Layer & Postgres

    Supabase Postgres with Row-Level Security from initial migration, least-privilege roles, atomic row locking, database check constraints, and pg_cron scheduling.

    • Postgres
    • Supabase
    • Row-Level Security (RLS)
    • pg_cron

  • Production Reliability

    Staging environments mirroring production, zero-downtime feature flags, dry-run validations, and deterministic rollback checkpoints. Automated alerting and call review loops.

    • Docker
    • Cloud Run
    • Vercel
    • GitHub Actions

  • Multimodal Processing

    Single-call invoice OCR parsing, automated image generation pipelines with local fallback models, audio transcription, and programmatic short-form video generation on rented GPU hardware.

    • Gemini Vision
    • FLUX ControlNet
    • LTX Video
    • ffmpeg

  • Python & TypeScript

    Production implementations in Python (scraping pipelines, valuation models, vector indexing) and TypeScript (Edge Functions, administrative dashboards, SDK automation). Software engineering fundamentals from mentored EPAM program (2022).

    • Python
    • TypeScript
    • Deno
    • Next.js

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
  • ElevenLabs KB in RAG mode
  • SQLite
Orchestration and agents
  • n8n workflow-as-code
  • @n8n/workflow-sdk
  • n8n MCP
  • agent teams of the agentic AI developer tool
  • 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
English: written working language with the US team (specs, reports, correspondence). Ukrainian (native), Russian.
Education
Master's degree, Department of Electronic Systems, Donbas State Technical University (2008-2013). Front-End Engineering Program with mentor, EPAM UpSkill and Strategeast (Jun-Nov 2022, Diploma ID 913874).
Code Verification
Commercial production code is protected under a mutual NDA; GitHub repositories remain private. For interview technical verification, I offer deep architectural walkthroughs of live production systems, live demos of proprietary products, or an interactive live build with an agentic AI developer tool.