Automation CV · real systems, not mockups
AI Automation Engineer — n8n · Claude · Telegram
Vienna, Austria · works in English · remote EU/US
I design, build and run AI agents and n8n workflows. Everything below is running in production on my own server right now — the same stack I deploy for clients. Third-party data is removed; the architecture, guardrails and results are real.
Pulled from my n8n instance every minute
How I build
Human in the loop
Agents draft, score and prepare. Anything irreversible — sending, paying, publishing — waits for one tap from a human.
Monitored, not hoped
Every system ships with health checks that verify results were actually written, plus failover and alerts.
Your data stays yours
Self-hosted n8n on an EU server, secrets in a credential vault, no client data in prompts or public repos.
01Lead qualification
- Problem
- Opportunities arrive as dozens of alert emails a day across nine platforms. Reading each one to decide 'worth it or not' eats hours and good leads still slip through.
- What I built
- An agent that collects every alert, opens and reads the full listing, scores it against a profile stored in a knowledge base and sends a decision card with a ready-to-send draft in the client's language.
- Stack
- n8n · Claude · IMAP · RSS · n8n Data Tables · Telegram Bot API
How it flows
- 1IMAP inboxes + RSS
- 2Dedupe & parse link, budget, rate type
- 3Read the full listing page
- 4Claude scores vs. profile (0–10)
- 5Telegram card: Take · Rewrite · Clarify · Pass
- 6Decision logged to data table
Results
- Scores 24/7 on a 15-minute cycle
- Reading the full page turned a '5/10, ask them' into a correct '2/10, pass' on a real listing
- Duplicates impossible by design (upsert on message ID)
- Rules live in a knowledge base — change behaviour without touching the workflow
Guardrail
Nothing is sent without a human tap; max 4 evaluations per run to cap cost and noise.
02Inbox intelligence
- Problem
- Replies and interview invites got buried among newsletters and notifications — and arrived without any preparation.
- What I built
- A branch on the same mail triggers that detects real replies from nine platforms, filters out newsletters, and for each one prepares a summary, priority, a draft answer and the three most likely interview questions with hints.
- Stack
- n8n · IMAP · LLM router · Telegram
How it flows
- 1Shared IMAP triggers (no double reads)
- 2Sender & intent classification
- 3Newsletter filter
- 4LLM: summary · priority · draft · 3 questions
- 5Telegram briefing
Results
- Answers drafted in the language of the incoming email
- Offline test: 3 of 3 real replies caught, 2 of 2 newsletters filtered
- Zero extra mail-server load — reuses existing triggers
Guardrail
Drafts only; replying stays a human decision.
03AI agent
- Problem
- Email, calendar, notes and a fleet of automations lived in separate tabs. Checking status or delegating a task meant opening five tools.
- What I built
- A conversational agent in Telegram with tools for mail, calendar, sheets, long-term memory, a knowledge base and even building new n8n workflows from a chat. Voice in and voice out.
- Stack
- n8n LangChain Agent · Claude · Whisper · Gmail & Calendar · Google Sheets · Telegram
How it flows
- 1Telegram message or voice note
- 2Owner verification
- 3Live system status injected server-side
- 4Claude agent picks tools
- 5Confirm step for irreversible actions
- 6Reply as text or voice
Results
- 15+ tools behind one chat
- Builds and deploys a new mini-agent from a conversation
- Tagged long-term memory it searches before answering
- Real-time status injection removed hallucinated 'system offline' replies
Guardrail
Only the owner can talk to it; money, deletion and system changes always ask first.
04Document processing
Client details confidential- Problem
- A client's back office retyped standardised multi-page financial acts from scanned PDFs into a spreadsheet by hand — slow and error-prone.
- What I built
- Drop a PDF or a ZIP of PDFs into a Telegram bot. Claude reads every page, extracts the fields, sums totals across all acts inside a file and appends one clean row per document to an Excel file in the cloud.
- Stack
- n8n · Claude vision · Google Drive · Excel (xlsx) · Telegram
How it flows
- 1PDF / ZIP to Telegram bot
- 2Unzip & queue, one file at a time
- 3Claude vision extraction
- 4Validate & sum across acts
- 5Append row to Excel in Drive
- 6Summary back in chat
Results
- One document = one row, totals summed automatically
- Batch uploads via ZIP
- Rate-limit safe: throttling plus 7 automatic retries
- Numbers written as values, never broken formulas
Guardrail
Client, company and document contents are confidential — shown here as architecture only.
05Reliability layer
- Problem
- Automations fail silently: a run can report 'success' while writing nothing, an API key dies overnight, a saved change never goes live.
- What I built
- A reliability layer over the whole system: a guardian that checks outputs were actually written, a publish guard, a 3-tier LLM fallback and an 08:30 briefing with the last 24 hours of errors.
- Stack
- n8n API · Anthropic · OpenRouter · Telegram · Docker on Hetzner (EU)
How it flows
- 1Every 30 min: compare live vs. saved versions
- 2Re-publish drifted workflows
- 3Check freshness of stored results
- 4LLM: Anthropic → OpenRouter → free model
- 5Telegram alert + morning briefing
Results
- First briefing surfaced 47 errors/day hidden in one workflow
- Guardian caught 4 silent failures that all reported 'success'
- Failover verified by deliberately breaking a key
Guardrail
Alerts are de-duplicated so real problems are not drowned in noise.
Toolbox
Automation
- n8n (self-hosted)
- Webhooks & REST APIs
- IMAP / Gmail
- Google Sheets & Drive
- Data Tables
AI
- Claude (Anthropic)
- OpenRouter
- LangChain agents
- Tool calling
- Whisper / voice
Delivery
- Telegram bots
- Next.js dashboards
- Docker & nginx
- Hetzner (EU)
- GitHub Actions CI/CD
Tell me which manual process eats your week. In 30 minutes you get a concrete plan and a fixed price.