Automation & AI Systems Engineer · open to full-time & freelance, remote only

I build systems that keep running after I've moved on.

Production automations across trading execution, CRM lead capture, AI content publishing, market monitoring, real-time translation, data aggregation. Every one is deployed, monitored and self-healing. Not a sandbox demo: these have real uptime, real failure modes, and real fixes behind them.

Hire me View systems
24/7 self-healing 1,249 signals backtested -70% inference cost 15s -> 3s voice latency
Vitalii Maslii, automation and AI systems engineer
Vitalii MasliiUkraine · EET
watchdogprocess auto-recovered
cost / request$0.31 -> $0.09
watchdog: ping executor -> 200 ok
What I automate

Engineering for production-grade automation

I take systems from architecture to live deployment — with monitoring, alerting and documentation included.

WF

Workflow Automation

n8n at expert level: custom nodes, SDK, MCP integrations, multi-branch pipelines with human-in-the-loop approvals.

AI

AI Agent Systems

Multi-model orchestration (Claude, GPT, Gemini), vision and voice pipelines, prompt engineering and cost optimisation.

CRM

CRM & RevOps

GoHighLevel from zero to production: missed-call text-back, nurture sequences, booking, webhook lead capture.

ETL

Data Pipelines

Multi-source ingestion with AI relevance scoring, deduplication and reliable scheduling that survives a sleeping machine.

Deep dives

Built different: reliability first

Self-healing production services

My systems don't just deploy — they survive. Every service ships with health checks, watchdogs and alerting.

  • macOS launchd services with health-check endpoints
  • Watchdog monitoring via n8n + Telegram alerts
  • Automatic process recovery after crashes
$ launchctl list | grep executor
com.p3.trade-executor  running
$ curl localhost:8787/health
200 OK · uptime 41d
# watchdog: auto-restart enabled
triggerAnalyse 42 new listing photos
gemini flashPre-screen done · 38 kept · $0.02
claude sonnetFinal scores ready · $0.07
resultcost $0.31 -> $0.09 (-70%)

Multi-model AI, optimised for cost

Hybrid pipelines route work to the cheapest model that can do it well — without sacrificing accuracy.

  • Hybrid Gemini Flash + Claude Sonnet vision pipeline
  • Roughly 70% lower inference cost per request
  • Prompt engineering, fallbacks and cost monitoring

Real-time voice and WebSocket systems

Bidirectional voice translation for live calls across three AI voice APIs — engineered for latency.

  • Alibaba Qwen + OpenAI Realtime + Google Gemini Live
  • Latency reduced from 15 seconds to 3
  • Real-time artifact elimination and voice-cloning fixes
Engineering stack

Tools I ship with

Grouped the way I use them: orchestration, intelligence, and the infrastructure underneath.

Automation platforms

n8n · expertcustom nodes / SDKMCP integrationsGoHighLevel CRMwebhooks

AI / LLM integration

Anthropic ClaudeOpenAI (GPT, Realtime)Google Geminiprompt engineeringmulti-model orchestrationAI agentsvision pipelines

Backend & DevOps

PythonJavaScriptFastAPIPostgreSQLSQLiteDocker / ComposeWebSocketlaunchdTailscaleGit
Integrations

APIs I've wired into production

Real integrations shipped in live systems — not logo wallpaper.

BinanceAlgo Order API · trading
TelegramBot API · Telethon userbot
GoHighLevelREST + MCP · CRM
LinkedInpublishing automation
OLX + DOM.RIAlisting ingestion
OSM GeoNominatim + Overpass
Zoomreal-time voice translation
X / Redditscraping + AI scoring
Systems

Case studies with measurable results

Self-initiated production systems, owned end to end — from architecture to monitoring.

status: live · self-healing
Trading · Python

P3 Trade Executor

Backtested a trend-following strategy (SHORT + SMA trend filter + risk management) against 1,249 historical signals before writing a line of execution code, then built the service that trades it: a Telegram signal listener on Telethon, an order-execution service on Binance's Algo Order API, and a watchdog that alerts me within ten minutes if the executor stops answering. It survives the failure mode that actually matters for something trading real money while you sleep — a silent process death.

OUTRuns as a self-restarting system service, not a background script. Recovery verified by killing it.
PythonBinance APIlaunchdn8nTelegram
0signals backtestedsource ->
status: live · 2 sources
Monitoring · FastAPI

"Tamara" Real-Estate Monitor

An end-to-end monitoring service that watches two listing sources, filters by real commute distance, and flags likely fakes — in a market with no listing API and inconsistent addresses. I built a self-hosted geocoding stack (Nominatim + Overpass on a custom OSM extract) after the public services proved too flaky, and a two-stage vision pipeline where a cheap model triages every photo and a stronger one only looks at what earns a second opinion.

OUTCut AI vision spend by roughly 70% with no change to what the end user saw.
PythonFastAPIPostgreSQLDockerGeminiClaudeOSM
-70%vision cost via hybrid AI pipelinesource ->
status: built · 3 engines
Real-time · Voice AI

Voice Translation for Zoom

Real-time two-way voice translation for video calls, built across three different realtime AI APIs as each proved better or worse at a different leg of the problem. Most of the engineering turned out to be audio, not AI: virtual routing, jitter buffers, voice-activity tuning, and a choppy-playback bug that traced back to Bluetooth silently switching codecs mid-call. Shipped with a web control panel.

OUTRound-trip translation latency down from roughly 15 seconds to 3.
WebSocket realtimeQwenOpenAIGemini LivePyAudio
15s -> 3saudio latencysource ->
status: live · 34 nodes
n8n · Content Ops

Content Factory

A 34-node n8n workflow that turns a topic dropped into Telegram into a published LinkedIn post. Claude drafts the copy in two languages against real copywriting frameworks, Gemini generates the image, and a human approves or sends it back for a rewrite with a single tap. Stateless by design — n8n's task-runner mode won't persist workflow state, so the state lives in the message thread itself.

OUTFull draft to publish loop, including a reject-and-regenerate cycle that never restarts the workflow.
n8nClaude APIGeminiLinkedIn APITelegram Bot API
0nodes in one orchestrated workflow
status: shipped · client-ready
RevOps · GoHighLevel

CRM Automation Suite

Production GoHighLevel workflows for a service business: a missed-call text-back flow that sends exactly one message to the customer — every follow-up and escalation stays internal, to staff — plus multi-step lead nurture, appointment booking with relative reminders, and webhook-based custom lead-capture forms. Where the platform's native form couldn't be restyled, I replaced it with a webhook trigger: full design control, same CRM automation underneath.

OUTCaught and redesigned a real logic flaw mid-build: customers were promised a callback 30 minutes before staff were told about it.
GoHighLevelWebhooksn8nREST APIMCP
0manual follow-ups required
status: live · scheduled
Data · Scraping

Multi-Source Aggregation Bots

Scrapers across X, Reddit, Telegram and niche job boards, with Claude scoring each item for relevance, intelligent deduplication, and scheduling that survives a laptop going to sleep. Several sources fought back — datacenter IPs blocked, undocumented APIs ignoring their own filter parameters — so each one needed its own approach rather than a single generic scraper.

OUTFive-plus sources collapsed into one deduplicated feed, delivered on schedule without manual runs.
PythonSeleniumClaudeSQLitelaunchdn8n
5+sources in one deduplicated feed
Readout

Numbers that are actually mine

0

historical signals backtested before any live order

0

cut in AI inference spend, same output quality

0

faster round-trip translation, 15s down to 3s

0

nodes in a single stateless production workflow

How I work

Automation in three steps

Self-directed, async-first, documented as I build.

01

Map the process

Audit the workflow, define triggers, success metrics and failure modes before writing a single node.

02

Build & integrate

n8n, REST APIs, LLMs and databases — tested in staging, with decisions documented along the way.

03

Deploy & monitor

Production deployment with health checks, watchdogs and Telegram alerting; iterate on real usage data.

FAQ

Questions recruiters actually ask

Something else? Reach out — I reply fast, async-friendly.

Contact me
Full-time or freelance?

Both — I'm open to a permanent full-time role or freelance/contract project work. The one non-negotiable is remote: async-first communication, documented decisions, regular written updates.

What's your timezone and overlap?

Ukraine — EET (UTC+2). I flexibly adjust overlap hours for US and EU teams.

What is your core stack?

n8n at expert level (custom nodes, SDK, MCP), Python / FastAPI / PostgreSQL / Docker, and production integrations with Anthropic Claude, OpenAI and Google Gemini APIs.

Do you only build prototypes?

No. Everything I ship runs in production: health-check endpoints, watchdog monitoring, self-healing restarts and alerting are part of my definition of done.

How do you handle incidents?

Systematically: silent-failure detection, log-driven debugging of distributed systems, post-incident fixes and documentation — the boring, reliable way.

Contact

Let's build your automation stack.

Open to full-time and freelance remote work in automation engineering, AI workflows, integration and RevOps engineering. One email — and I'll show you live systems, not slides.

vitaliymasliy100@gmail.com LinkedIn Telegram