AI Automation Developer

Automations tested before they go live.

Automations and AI assistants for agencies, built with n8n, Zapier and Claude. Planned with you first, tested before launch and handed over with a written guide.

Mock-up: made-up example, real checks in order

Mock-up: made-up accounts and figures

What goes wrong, and the fix

What clients usually run into with automations, and how I fix each one.

01Secured

The problem: Our automation was open to anyone who found the link.

Mock-up: made-up messages, real check results

The fix

Every workflow checked, incoming messages confirmed as genuine, and secrets kept out of workflows and logs.

23 workflows checked one by one, from the case study: Articles written and posted automatically

02Error free

The problem: It worked in the demo, then broke on real orders.

Illustration: real test results, simplified

The fix

Tested on real information, retested after every change, with an alert if anything slips through.

18 of 18 checks passed before release, from the case study: Campaign folders created automatically in SharePoint

03Production

The problem: Going live meant duplicate posts, downtime and no way back.

Mock-up: made-up blog names, real steps

The fix

Switched over in small steps with a 24-hour watch, and the old way kept as a backup.

13 blogs moved, every zap kept as a backup, from the case study: Blog publishing on n8n for SiteSwarm and Duda

04Optimized

The problem: A heavyweight setup for a small job, and a bill that kept growing.

Illustration: real costs, drawn to scale

The fix

Right-sized to the job: simple rules before AI, one small server, nothing extra.

US$12 to 14 a month to run the server, from the case study: Private n8n server on AWS, rebuildable from code

05Documented

The problem: Only the person who built it knows how it works.

Mock-up: real guide sections, drawn page

The fix

A how-to guide, a change log and handover notes, so your team can run it without me.

255 lines in one how-to guide, from the case study: Blog publishing on n8n for SiteSwarm and Duda

06Guarded AI

The problem: The AI made things up, or spent money nobody approved.

Mock-up: made-up documents, real 0.45 cut-off

The fix

Weak search matches never reach the AI, and a person approves anything that spends money.

0.45 minimum relevance score, from the case study: A Slack assistant that answers from your own documents
01 / 06

Selected work

Illustration of the final result

01Digital marketing agency, 13 client blogs

Blog publishing on n8n for SiteSwarm and Duda

A marketing agency writes its clients' blog articles with blogr.ai, an AI writing tool. These two n8n workflows publish each article to the right client's SiteSwarm or Duda website automatically, keep old links working, send the new page to Omega Indexer so search engines find it, and log every article.

13
blogs moved, 10 on SiteSwarm and 3 on Duda
18 of 18
old zaps switched off, none deleted

Mock-up: made-up conversation

02Software platform for dental clinics

Appointment chatbot for dental clinics

A chat assistant on each clinic's Facebook page. Patients book, cancel or check an appointment in English or Filipino, straight against the clinic's own appointment system, and staff can step into any conversation. It also sends reminders and follow-ups.

377
recorded test conversations
33 of 39
replies passed the AI judge

Illustration of the real setup, simplified

03Digital marketing agency

Private n8n server on AWS, rebuildable from code

A marketing agency's own private server for its n8n automations, on AWS for about US$12 to 14 a month. The whole setup is written as code, so it can be rebuilt exactly, every change is checked automatically, and the server backs itself up every night to a second location.

US$12 to 14
a month to run
30 days
of off-site backups, in a second region

Tools I use

SuccessErrorSuccessErrortruefalseSuccessErrorSuccessErrorSuccessErrortruefalse012SuccessErrorSuccessErrorSuccessErrorSuccessErrortruefalsetruefalseSuccessErrortruefalseFetch TopicsFetch TagsBuild Filter PayloadSlack TriggerNeeds Reranking?Extract ScoresSet QueryBuild Rerank PayloadRerank (Cohere)Parse PineconeResultsBuild AI Payload(Reranked)Build AI Payload(Pinecone)Embed QuerySearch PineconeSlack NormalizerOutput Parser:IntentSet Query MetadataQuery TaggerChat ModelOutput ParserSearch Pinecone(Fallback)Has Matches?LLM: Founder(gpt-5.2)LLM: Intent Router(gpt-5-mini)Route by ModeVA Info AgentChat ModelMemoryLLM: Query Tagger(gpt-4.1-mini)Send Slack ReplyEnd (No-Op)Founder Agent (WithContext)Chat ModelMemoryLLM: FounderFallback (gpt-5.2)Intent RouterChat ModelOutput ParserLLM: VA Agent(gpt-4.1-mini)Output Parser:Query TaggerSet ContextResponse NormalizerMemory: FounderMemory: VAEval TriggerFounder Agent (NoContext)(Deactivated)VA RAG Agent(Deactivated)Skip Rated RowsWrite Response toSheet(Deactivated)Eval NormalizerHas Thread?Fetch Thread HistoryFormat ThreadContextNo Thread(Passthrough)Dedup CheckIs Query Empty?Send Empty ReplyValidate ResponseLog Delivery Failure

My main automation tool

Most of what I build runs here: AI agents that answer from your own documents, connections between the tools you already use, and custom code where a ready-made step falls short. It runs on servers I set up, back up and watch.

  • AI agents
  • Webhooks and APIs
  • Custom code
  • Self-hosted on AWS
  • Failure alerts

Make

For teams already on Make

Scenarios with routers, filters and data stores. Before building, I compare Make with n8n and Zapier on price and fit, so the tool suits the job.

  • Routers and filters
  • Data stores
  • Webhooks and APIs

Zapier

Fixed, documented, then moved

Fixing and documenting zaps a team already depends on, and moving them to n8n when the task bill grows, with the old zaps kept as a backup.

  • Paths and filters
  • Code steps
  • Fixing live zaps
  • Moving to n8n

AI and search

  • Claude and Claude Code
  • OpenAI
  • Gemini
  • Pinecone
  • Cohere rerank
  • Vercel AI SDK
  • AssemblyAI
  • Leonardo.ai

Publishing and content

  • SiteSwarm
  • Duda
  • blogr.ai
  • Omega Indexer
  • WordPress
  • Blogger
  • Postiz
  • Facebook and X

Business tools and messaging

  • GoHighLevel
  • Jobber
  • Slack
  • Telegram
  • Facebook Messenger
  • ClickUp
  • SharePoint (Microsoft Graph)
  • Meta Marketing API

Google, email and notifications

  • Google Sheets, Docs and Drive
  • Google Calendar
  • Gmail API
  • Resend
  • Firebase Cloud Messaging

Databases and servers

  • Supabase
  • PostgreSQL
  • Redis and BullMQ
  • Terraform
  • AWS
  • Cloudflare Pages
  • Docker
  • Caddy
  • GitHub Actions
  • Healthchecks.io

Programming

  • TypeScript
  • Node.js and Express
  • Python
  • NestJS
  • React
  • Vue
  • Next.js
  • Electron
  • Playwright
  • ffmpeg

How I work

Before anything is built

  1. 01

    Understand

    • The business goal and who it serves
    • The tools, data and limits involved
    • What "done" looks like
    • What already exists
  2. 02

    Scope

    • Requirements written down
    • What is not included, stated
    • Open questions answered
    • A test plan and a way back

Nothing is built until scope is agreed

Building and shipping

  1. 03

    Build and test

    • Fixed test cases
    • Real data, not samples
  2. 04

    Go live

    • One client at a time
    • Old way kept as a backup
  3. 05

    Hand over

    • A how-to guide and a change log
    • Handover notes for the tester
  • 01

    Agree what we are building first

    What is included, what is not and any open questions are written down and agreed before building starts.

  • 02

    Test with real information

    Before launch, each automation runs on real examples from your business, not made-up ones, and the same checks run again after every change.

  • 03

    Keep the old way until the new one proves itself

    Work done by hand is handed over a bit at a time, with your team checking the results. An old automation is paused, not deleted, so you can switch back.

  • 04

    Let a person approve anything that spends money

    If an automation can spend money, it prepares the change and waits for a person to say yes.

  • 05

    Make quiet failures noisy

    Automations check in each time they run. If one fails or stops, you get an alert, instead of finding out weeks later.

About

I build automations and AI assistants for agencies and service businesses, using n8n, Zapier and Claude, and I set up the servers and checks that keep them running. I work at an AI virtual assistant agency.

Before that I spent several years building web apps and the systems behind them, in Java, NestJS, Angular, PostgreSQL and AWS Lambda. So I treat automations like software: every change is saved and tracked, tested, and handed over with a how-to guide.