I build AI into the parts of a business where it earns its keep, and I build it so it can run without a babysitter.
The system I am proudest of is a fully autonomous daily social publishing pipeline built for a public safety nonprofit, running entirely on Zapier. A schedule trigger fires. Claude writes the post copy plus an image prompt keyed to the day of the week. A three attempt looping retry handles failures. An image is generated through a webhook call and hosted on Cloudinary. A Python step running in code executes a Claude vision review of the image before anything is allowed through. A storage key enforces idempotency so nothing posts twice. A Google Sheet acts as both queue and audit trail. A second workflow publishes approved rows to Facebook and Instagram. It has been running unattended since June 2026.
What I build: content and reporting agents, document processing pipelines, research and summarisation workflows, AI assisted intake and triage, internal chat tools grounded in your own data, and AI review gates dropped inside automations you already have.
What I insist on: a human in the loop wherever the stakes justify it, idempotency keys, retry and fallback paths, spend controls, and logs a non engineer can read. Most AI projects do not fail on the model. They fail on operations.
Stack: Claude API, OpenAI API, Zapier and Make.com AI steps, Python, Google Apps Script, Cloudinary, webhooks.