Lead Qualifier — score scraped leads against your own ICP

You already have a scraper that produces companies or contacts. This scores each one against your ideal customer profile and hands back a sorted list:

score	0–10 fit, so you can sort and work top-down
verdict	qualified · review · disqualified
reason	one label: strong_icp_match, wrong_geography, excluded_account, no_authority, …
signals	the per-signal probabilities behind the score, so you can re-threshold without paying again

Point it at any dataset — LinkedIn, Apollo, Google Maps, a company-list scraper, your own CSV — describe who you sell to, and run it.

Why not just filter by keyword

Keyword rules break on the cases that matter. This run, on a corpus of 240 labelled leads, got all of these right:

• Lenze SE, lenze, and Lenze Drive Technologies all recognised as the same excluded competitor — 0.988 accuracy on exclusion matching
• Vestlund Nordic AB recognised as a subsidiary of an excluded customer
• a company in the UK correctly held for review rather than killed, when the target regions were Germany, Austria, Switzerland, the Netherlands and the Nordics — neighbouring markets are a judgement call, not a reject
• process automation held for review against a target list of

  industrial automation
  — adjacent, not unrelated
• a perfect-fit company whose only contact was info@ sent to review, because there is no decision maker to call

It is built not to lose your leads

The expensive mistake is throwing away a real lead, so the whole thing is asymmetric by design. Measured on the labelled corpus:

Verdict accuracy, three buckets	0.929
Qualified leads wrongly disqualified	0
Disqualified leads wrongly passed as qualified	0
Precision of the qualified bucket	0.833
Recall of the qualified bucket	0.875

Every mistake it does make is one step in the cautious direction — into review, where a human sees it. Uncertainty always widens the net:

• a sparse row is never disqualified, because thin data is not evidence
• a low-confidence judgement is downgraded to review, in both directions
• an account showing an active buying trigger is never hard-killed on a weak fit
• the only thing that may disqualify outright is your own instruction: your exclusion list, or a confidently unrelated industry or region

What you pay

You are billed per lead scored. Not per lead returned, and never for a lead that could not be judged — those come back flagged with billed: false. maxItems is a hard ceiling on the run, so the cost is known before you start.

If more than a fifth of the leads cannot be judged, the run fails instead of handing you a list that looks complete but is not. You are not billed for them.

Input

• buyerProfile (required) — who you sell to. target_industries, target_size, target_regions, buyer_roles, buying_triggers, and exclusions with competitors, existing_customers, blocked_domains. Every field optional; more detail means sharper scoring.
• sourceDatasetId — any scraper run's dataset. Or paste records into leads instead.
• keepVerdicts / minScore — what lands in the output. Everything is scored and billed regardless; these only trim the list you get back.
• maxItems, concurrency, idField, includeAnswers.

Output

One row per returned lead, with your original fields kept alongside score, verdict, reason, confidence, rationale, rules_applied and optionally signals. A RUN_SUMMARY record in the key-value store holds the counts, timing and which verdicts were kept.

Speed

About 13 leads a second at the default concurrency; 1,000 leads in roughly a minute and a half.