Joshua P.

An AI agent near your inbox raises one question before any other: what stops it sending something you would not have sent? I build the approval step as workflow structure, so anything that writes or sends renders as a draft and waits for a yes. That is the same instinct I bring to everything else I build. The failure I design against is silence: a workflow that ran clean for a month stops at 2am, and the first anyone hears about it is a customer complaining a week later. So I register an error workflow, route failures to a channel someone actually reads, make writes idempotent so a re-run doesn't create a second copy of the same record, and back off retries instead of hammering an API until it locks you out. What I build: - n8n workflows and API integrations: webhooks, CRM writes, scheduled jobs, and the error lane that tells you when one stops. - AI agents and LLM steps inside n8n (Claude, OpenAI or Gemini), with structured outputs and a human checkpoint anywhere the agent would otherwise act on its own. - WhatsApp and Telegram automation: quoting, booking and support flows, including permissioned bot commands that leave one audit row per action. - Data pipelines and scraping: Python ETL, Apify and Playwright sourcing, deduplication, scoring, and delivery into Airtable, Sheets or your CRM. When an LLM does the judging, I audit what it decided rather than trusting the score. On my last scoring build an automated spot-check flagged two rows; I checked both against the source: one was a real misgrade and was moved out before delivery, the other was a false alarm. That check is part of the work, not an extra. I have completed Anthropic's Claude 101 and Claude Code 101 courses and work in Claude Code daily. How it works: send me the problem and you will hear back the same day, usually within a few hours, with either a fixed-price quote or the two questions I need in order to give you one. Builds are scoped deliverables with handover docs, so your team can maintain and change the system after handover. If you do not yet know which process to automate, say that and we will scope it on a call before you commit to anything.