Astragal

Astragal serves a Head of Collections at a bank, NBFC or insurer of over a thousand people whose voice pilot only ever called whitelisted numbers. Before go-live, someone must certify in writing that the first ten thousand numbers may lawfully be dialed by an artificial voice.

Astragal is a versioned REST API that answers clear, hold or refuse for every number on a dial list. Each number runs a fixed, published rule ladder that stops at the first refusal.

A clear carries a decision id and an expiry for the dialer to present.

Rules cannot judge whether consent evidence covers a call, so in Astragal a fine-tuned text classifier does. Rules resolve a structured Digital Consent Acquisition reference; the rest, US consents included, is free text from four systems.

The classifier reads each wording with its channel and proposes what it covers, including any artificial-voice disclosure. A smaller embedding model matches one person's records across them.

No model in Astragal can clear or refuse a number, only hold it. A proposed verdict applies above a fixed threshold once a lender's compliance reviewer confirms it, one wording at a time; below it, the numbers hold.

Each confirmation adds a wording and a human verdict to a coverage corpus we own, pooled across lenders without numbers or identifiers, which trains the classifier.

Astragal's heaviest compute is the onboarding backfill, which classifies a new lender's whole consent history in one pass. At a mid-size lender that means single-digit millions of records, unbilled.

A republished rule pack re-evaluates every stored record for every customer; between those, a lender is nearly idle. Dial lists land before the morning and evening calling windows, up to 50,000 rows a batch, answered within four minutes.

Consent wording holds the exact promise a lender made to its customer, so Astragal sends none of it to an inference endpoint another company operates. Everything runs on cloud infrastructure in Columbus, Ohio, with no second location and no in-country deployment, so an Indian lender's numbers are checked in the United States.

Astragal moves both models from on-demand to reserved GPU capacity in two dated steps. Through Q1 2027 they run on on-demand cloud GPU capacity we control in Columbus, Ohio.

From Q2 2027 the coverage classifier moves onto an open-weight model, fine-tuned and served from reserved capacity there, sized for rule pack republication, with backfills still on demand. The consent-matching embedding index follows in Q4 2027.

Astragal S.A.S. is eight months old and its API is live: founded in January 2026, incorporated in Medellín, Colombia. Its three founders came to this problem from collections, from the pilot's engineering and from six years in consent operations.