account-dossier-agent turns a company's website, blog, job board, SEC filings and news coverage into a cited research dossier, then drafts outreach that can only reference what the dossier found.
A five-stage account research pipeline. Scout pulls a company's site, blog, job board, SEC filings and news. Extract requires a verbatim quote for every claim. Verify checks each claim against how trustworthy its source is. Render writes a footnoted dossier, and Outreach drafts copy that can only cite what the dossier has already verified.
Account research usually means a rep skimming a homepage for five minutes, or an LLM inventing a number nobody checked. This gives every outbound email a sourced fact to open with. When a company has little public information, the dossier says so, which beats a confident paragraph that falls apart on the first reply.
Fetches the company's site, blog RSS, job board, SEC filings and news. It respects robots.txt, rate-limits itself and uses a self-healing cache.
One LLM call per source. Every claim needs a verbatim quote, and if the quote isn't an exact match in the source, the claim is dropped.
A claim backed by two or more sources, or by one Tier 1 or Tier 2 source, is marked verified. A claim with a single Tier 3 source is flagged for review.
A markdown dossier with every sentence footnoted [Cn] back to its source URL. If there's little public data, the dossier is short.
Drafts email and LinkedIn copy that can only cite claims the dossier has verified, checked by code.
All free, none need an API key.
Homepage, about, product, pricing and careers pages, plus whatever the company's own blog or newsroom feed publishes, found automatically.
Detects Greenhouse, Lever or Ashby on the careers page and pulls exact role, department and location data from their public APIs, so there's no scraping or model guesswork involved.
Looks up the company's CIK and pulls its most recent filings, preferring short current reports (8-K/6-K) over long annual reports so the model reads recent, dense information instead of boilerplate.
Queries the open GDELT news index for recent coverage. This is the lowest-trust tier on purpose: a single news hit is flagged and never treated as verified on its own.
Every claim carries a verbatim quote from its source, checked in code. If the model can't quote it exactly, the claim is rejected, and the rejection shows up in the output.
A company with little public information gets a dossier that says “limited public footprint” instead of a padded paragraph of invented details.
Quote matching, citation checking and pain-signal matching are plain code. A model is never asked to grade its own work.
If extraction crashes after a source has been fetched, the next run notices nothing was extracted from it and tries again, rather than marking it done.
[C93] Acme Retail is hiring 2 AP clerks to join its finance team.
“Hi AP Manager, I came across that Acme Retail is hiring 2 AP clerks to join its finance team [C93]. We help finance teams scale without scaling headcount by automating the full accounts-payable cycle. Would you like to learn more?”
Passed. Every citation resolves to an allowed claim, and no uncited sentence contains a number, $ or %.