Systems we built and run in production — not slide decks.

Six automations from the last two months: a voice AI receptionist taking real calls, a lead funnel built from zero to production in days, a deduplicated multi-source lead pipeline, a multilingual WhatsApp agent, the remote ops bridge that runs all of it from a phone, and a lead-sourcing pipeline that works around Facebook having no public API.

ElevenLabs Conversational AI
agent config + call log view
(screenshot pending — see report)
01 — Voice AI
Katerina — voice AI reception
Problem
Small businesses lose leads and admin time to calls nobody's free to answer — hiring reception staff for the call volume doesn't pencil out.
Solution
A production ElevenLabs conversational voice agent handling real inbound and outbound calls, wired into Google Calendar and Gmail for scheduling and follow-up, Notion for task memory and per-caller personalization, and native DTMF touch-tone support so it can navigate IVR menus on its own. The same architecture runs Arseniy's own admin work internally — booking, email triage, weekly self-review — before being deployed to clients.
Stack
ElevenLabs Conversational AIn8nGoogle Calendar APIGmail APINotion APIZadarma / SIP
Result
Live in production, handling real client calls — first paying client onboarded on a dedicated persona.
Airahome landing page — emergency home repair callout, Algarve
02 — Lead Funnel
Airahome — 4-language repair callout funnel
Problem
Expats and foreign property owners in the Algarve need home repairs fast, in their own language, without hunting for a contractor — and every minute of delay after they submit a request loses trust.
Solution
Built aira-ai.net/airahome from zero to production in days: a 4-language landing page (EN/DE/FR/PT) with tiered urgency-based pricing, a lead form and WhatsApp entry point wired straight into n8n. One webhook pushes every lead to Telegram and immediately triggers an outbound ElevenLabs voice callback in the customer's own language confirming the request. A second webhook powers the on-page chat widget. No manual dispatching in the loop.
Stack
Static HTML/CSS/JSCloudflare Pagesn8n webhooksElevenLabs outbound callingClaude (chat + intake)
Result
Full lead-to-callback funnel live in production, zero to shipped in days.
n8n workflow canvas
fb-notify-relay
(screenshot pending — see report)
03 — Lead Monitoring
Multi-source lead pipeline with dedup
Problem
Handyman leads surface in Facebook groups and Telegram expat groups, in four languages, faster than anyone can manually watch — and scraping Facebook directly with automation risks getting the account banned.
Solution
Two independent, ban-safe intake sources feed one pipeline: a Telethon-based Telegram userbot reading expat group chats from a real logged-in account (no scraping, no automation flags), and a push-based Facebook relay through n8n. Both land on a single intake listener that keyword-filters across EN/DE/FR/RU, dedupes by content hash in SQLite, and fuzzy-matches near-duplicate leads from different sources into one card before pushing it to Telegram. A read-only stats dashboard tracks demand across two independent business verticals from the same data.
Stack
Python / TelethonSQLiten8nCloudflare Tunnel
Result
Continuous deduplicated lead flow, live dashboard, replacing a person reading group chats all day.
WhatsApp bot conversation
(Baileys + Claude Haiku)
(screenshot pending — see report)
04 — WhatsApp Agent
Sofia — multilingual WhatsApp agent
Problem
One WhatsApp number needed to do two different jobs — cold-outreach hotels for B2B sales, and handle inbound customer repair intake — each in the customer's own language, without maintaining two separate bots.
Solution
A single n8n workflow behind a Baileys WhatsApp connection dynamically builds a different system prompt per context: a structured intake persona that asks for the problem, address and urgency, quotes tiered pricing, and emits a hidden structured lead tag for downstream automation — or a B2B outreach persona per hotel and city. Same agent, same deployment, branching logic decides the job.
Stack
Baileys (WhatsApp Web protocol)Dockern8nClaude Haiku 4.5
Result
One bot covering both outbound sales and inbound intake — one thing to maintain instead of two.
Telegram → n8n → VPS
remote command flow
(screenshot pending — see report)
05 — Remote Ops
Remote ops bridge — infra from a phone
Problem
Changing an n8n workflow or a VPS config normally means sitting at a laptop with SSH open — not practical mid-commute or away from a desk.
Solution
A Telegram message to the ops bot gets parsed by Claude into an intended command, then routed either to a laptop tunnel if it's online, or queued directly on the VPS through a local command server — results report back over Telegram either way. A companion "Code Bridge" workflow lets a running agent session pause on a genuine decision point, write a waiting task into Notion, ping Telegram, and resume once answered — instead of hanging silently on a blocking question. A mandatory backup-before-change step is built into the workflow rules themselves.
Stack
n8nTelegram Bot APIClaudeNotion APIPython command serverCloudflare Tunnel
Result
Infra changes and agent decisions approved from a phone, with a built-in safety net against silent hangs and unbacked-up changes.
Anonymous scraper + Android deep-loop
dual lead-source pipeline
(screenshot pending — see report)
06 — Lead Sourcing
Working around Facebook having no API
Problem
Facebook has no public API for reading group content — and that's exactly where local repair and window-installation leads live, in closed expat groups.
Solution
Built two independent systems running in parallel. An anonymous HTTP scraper reverse-engineers the logged-out mobile site's undocumented, non-deterministic behavior — the same request sometimes returns a hydrated page with posts and author IDs, sometimes an empty JS shell, with no dependency on IP or timing — and retries with a clean session rather than giving up on the cycle. In parallel, a device-level automation loop (MacroDroid on a real Android phone) emulates a human scrolling the group list and reading text off the screen, sidestepping the API restriction entirely. Both feed the same deduplicated, classified intake pipeline used for every other lead source.
Stack
Python (HTTP scraper)MacroDroid (Android)n8n intake relayPython + SQLite classifier
Result
A working lead stream out of closed Facebook groups — zero official API keys, zero account-ban risk.

Want something like this built for your business?