TasedevCaller

Tasedev Caller runs outbound cold-call campaigns with no human dialing. A tenant selects a slice of a shared national directory by industry, state, rating and review count, or uploads their own CSV or XLSX. A worker walks the leads one at a time: it fetches each lead's website, has a language model extract concrete, falsifiable weaknesses from what it actually saw, maps those to services the tenant sells, writes an opener citing the finding, and places the call.

The agent discloses it is AI on the first turn, qualifies the prospect, captures an email, books a calendar slot and honours an opt-out immediately. Outcomes return by webhook and reconcile against the call record. A manual track emails the same directory instead of calling it.

The entity model has two halves that deliberately do not mix. The tenant-owned half is tenants, members, campaigns, leads, calls and meetings, with usage and billing events beside them. The shared half is one canonical row per business, list-partitioned across the fifty states plus the District of Columbia, a cached website analysis keyed by domain, and a global do-not-call list. Tenancy is enforced in three layers: row-level security on every tenant-owned table, a shared directory reachable only through the service role, and a per-tenant audit table over the shared rows. Built solo, schema to dialer, in forty days.

This is where an unreliable generative step is made safe enough to speak aloud to a stranger. One HTTP fetch pulls the lead's homepage and extracts cheap deterministic signals: title, meta description, viewport tag, scheme, response time, and visible text capped at 80 KB of HTML and 3,500 characters.

Those signals, the directory record and the tenant's own service catalog go to the model under a prompt that permits only a closed set of fifteen weakness tags and demands one sentence of falsifiable evidence per tag, with worked examples of what passes ("no viewport meta tag was found") and what does not ("mobile experience could be improved"). Anything outside the tag set is dropped on the way back, and recommended services are intersected against the tenant's catalog, so the model cannot invent a weakness the system has no vocabulary for or sell a service the tenant does not offer.

A second step turns the surviving evidence into spoken copy with the business name and city left as placeholders, so it caches for seven days per tenant, industry and service and personalises per lead at dial time. Analyses cache ninety days per domain, so ten tenants pitching one business pay for one analysis. The closed tag vocabulary is what makes the rest work: it turns free text into something a scoring function, a cache key and a service mapping can all be built on, and it is why a failed fetch degrades to a weaker pitch instead of a fabricated one.

ARCH