Turning messy business processes into agents

Most "AI projects" I see start from the model. I start from the workflow.

Before any code, I look for a business process that's slow, manual, and repeated — a sales team hunting RFPs across dozens of procurement portals by hand, an expert aligning products to standards one PDF at a time. That friction is the real spec. The model is just the tool that removes it.

The pattern that keeps working

Find the human steps. Wrap the deterministic parts in real code. Then let an LLM agent orchestrate and judge — not write the scrapers. Structured output to keep it predictable, a tight tool loop to keep it cheap, and a human in the loop where the world fights back (bot walls, ambiguous data).

What I care about isn't "an agent." It's the business result on the other side.

4× more qualified opportunities. Half the time to match a product to a requirement. A proposal drafted in minutes instead of days. Start from the mess, ship the system, measure the outcome.

More notes to come as I build.