Mathew Joseph

Developers ask AI coding assistants to set up their deployments, and leave quietly when it breaks. I proved with a reproducible test that the file those assistants read failed its own spec, then specified, directed and shipped the fix and an official toolkit, executing through an AI coding agent. Tradeoff: directing an agent made it fast, so a second reviewer approved every change before merge.

In my third week I wrote the standard for answering interested prospects: answer directly, give one piece of proof, offer a low-friction next step. When volume outgrew human reading, I designed a classifier that checks each conversation against the ICP and records every decision. Tradeoff: it reads the whole thread, not reply 1, at extra cost, as a contact can sound keen and then close the door.

One practitioner's LinkedIn audience was converting to meetings noticeably better than the rest, and nobody knew why. I turned it into a model of what makes a source worth monitoring, and set up a record linking meetings to their source campaign. Within a month it showed my top-ranked source had booked none, so I reversed the model. Tradeoff: a model can fail where a list cannot; I chose that.

With two separately built GTM frameworks in play, I moved 100+ GTM documents into version control in one operation, added a source-of-truth register and an append-only decision log, then merged both frameworks into one customer definition. Tradeoff: I accepted a reversal of my own design, screening on company data before the costly signal work, because screening cheaply first was better.