Web-Presence Engine — lead-finding to audit to outreach

This is the most commercial of the builds, and the one that runs an agentic pipeline end to end: find candidate small businesses in an area, work out what web presence each actually has, decide whether it is worth rebuilding, generate a prospect-facing audit and a mockup, and hand a shortlist to me to send. It is a real side project, not a toy, which is what makes it a useful test of what agents can and cannot be trusted to do in a workflow with money at the end of it.

The pipeline

Discovery finds businesses at volume. A scoring agent then verifies each one's actual web presence and makes the call the whole thing turns on: keep this as a rebuild lead, or drop it. A thin, dated, one-page, or third-party-templated site counts as a lead, because that is exactly the segment worth rebuilding, so the expensive mistake is quietly discarding those by calling them good enough. From the qualified leads it can generate an audit that leads with the business consequence rather than the metric, and a disposable mockup to paste into a first message.

What it taught me about trusting agents with judgement

Two things from this build generalise well beyond it.

The first is that model choice on the scoring step is not a vibe. I held two models against a hand-labelled gold set and found they agreed on the easy majority and diverged on exactly the boundary the task lives on. The cheaper model quietly dropped real leads by calling thin sites good, and at volume that gap was large. Most of the difference turned out to be prompt adherence rather than raw capability, which I only knew because I controlled for the prompt instead of confounding it with the model. The gold set also cleaned its own ground truth as a side effect, surfacing a handful of my own labels that were wrong.

The second is the send-gate. Everything the pipeline produces is a draft. Nothing reaches a real business without me. A message sent to a real person cannot be unsent, so that step is deliberately the one place the automation stops and a human decides, regardless of how confident the agent is. The rest of the pipeline runs autonomously precisely because that irreversible edge is fenced off with a human on it.