n8n workflows, RAG pipelines on pgvector, and LLM integrations for teams that have outgrown copy-paste and a folder of Zaps. Built to be handed over, monitored, and maintained.
I design and ship production automation: n8n workflows that run unattended, retrieval pipelines that keep an LLM grounded in a team's own data, and integrations that stitch Supabase, Slack, and third-party APIs into one path.
The reason those systems hold up is the route I took to them. I trained as a mechanical engineer, wrote software, then spent years inside executive operations — running the calendars, inboxes, and handoffs I now automate. I have been the person on the receiving end of a broken workflow, so I build for the Tuesday morning when something fails and someone who is not a developer has to understand why.
I write about the work too, mostly practical guides on self-hosting and integration. Recent pieces are below.
Grouped by what it does, not by how well I claim to know it.
A social intelligence layer for teams: it reads what is happening across social channels and exposes that signal to agents over MCP, so an assistant can act on it rather than just summarise it.
An agent that works an inbox: LangGraph drives the decision loop, the Gmail API supplies the mail, and a custom MCP server gives the model the tools to sort, label, and draft.
// client automation work is under NDA — case studies in progress
// earlierWorkflows that run unattended — triggers, error handling, retries, and a handover doc your team can actually follow.
Retrieval pipelines on pgvector and Supabase, so a model answers from your documents instead of guessing.
Mapping the manual process first, then removing the parts a person should never have been doing by hand.
Practical guides, written while working through the problem.
Getting a durable n8n instance running on your own infrastructure, from container to first workflow.
What an API actually is, and how to call one, for people who keep nodding along in meetings.
A look at where the tooling was heading, written at the end of 2024.
The manifest, the moving parts, and a small extension built from scratch.
Loading, reshaping, and inspecting data — with a Plotly bonus at the end.
What the tutorials leave out about finishing something and putting it in front of people.
Two ways to get a model working on your rows: the native
=AI() formula, and a custom GPT function in Apps
Script when you hit its limits.
What I walked colleagues through: triggers and actions, branching, Tables and Sub-Zaps, AI nodes — and when to reach for n8n instead.
Desktop scheduled tasks versus Claude Code on the web: which one runs with your laptop closed, and which one can reach your local files.
Describe the process and where it currently breaks. I will reply with whether it is worth automating and roughly what it takes.