Statewide sweep
- Discover0% — not startednot started
- Dedup0% — not startednot started
- Reconcile0% — not startednot started
- AI ReviewAI — Claude judges each lead the fuzzy pass left undecided against its best DB candidate. Confident verdicts apply themselves; the rest go to the review queue.0% — not startednot started
- Enrich0% — not startednot started
- Emails0% — not startednot started
- VerifyAI — Claude judges whether each practice really exists, from NPI, Apify status, a stealth Maps read and the site.0% — not startednot started
- AI FitAI — Claude rates each lead against the Glidewell product catalog. Stamps ai_fit_* alongside the metric tier without overwriting it.0% — not startednot started
Re-run a single stage
One run takes every selected state from discovery through AI fit. Discovery calls Apify (billed per place) and scrapes the states in parallel; each state then goes through dedup → reconcile → AI review → enrich → emails → verify → AI fit on its own. Emails checks each practice's homepage and contact pages, and fills the export's Email column. Percentages come from what each stage actually reports — hover a tile to see what it is measuring. A stage badged AI is judged by Claude and costs tokens per lead — AI review, Verify and AI fit, all three running on Amazon Bedrock. Every other stage is deterministic. If Bedrock cannot be reached the AI stages fail and say so, and the rest of the run still finishes — as it does when the email harvest fails. A stage at 100% is a button: click it to show what it produced below — all scraped, everything, the review queue, the verdicts — and net-new stays one chip away.