An inbound call hits the voice agent and runs through a qualification state machine: greet, qualify, book, or route urgent work to a human. On the outbound side a dispatch layer enforces calling windows, attempt caps, and backoff per lead. Bookings land in the CRM with job details, and the payment flow closes the loop. It runs on AWS behind Cloudflare. The open core-dispatch engine on GitHub is that same qualify-and-schedule state machine, cleaned up and dependency-free, with its own tests and CI.

A parser normalizes VAPI, Retell, Twilio, or plain-text transcripts into one call model. An outcome classifier separates humans from voicemail and IVR, a signal detector flags objections, pricing talk, bookings, and do-not-call requests with the exact lines as evidence, and a deterministic scorer grades the conversation. Pure Python standard library, zero runtime dependencies, fully offline — it runs the same on a laptop as it does in the pipeline.

Ingestion workers pull ads from public ad libraries, a normalizer maps every source into one canonical ad model, and the analysis layer detects hooks and angles — question hooks, stat hooks, urgency, social proof — carrying the matched text as evidence, then scores the creative. A Next.js dashboard sits on top. The open adoracle repo carries the analysis engine: normalizer, detectors, and scoring.

A ranking model orders the dormant list by win-back potential, and a scheduler builds compliant call queues: timezone calling windows, quiet hours, attempt caps, cooldowns. The voice agent works through the queue and outcomes feed straight back — opt-outs suppress permanently, voicemails back off, interest routes to booking. A readiness gate refuses to start a campaign unless config, suppression list, and windows all validate.

Discovery and enrichment feed a scoring model that grades audit fit with an explainable factor breakdown. Every lead then passes a fail-closed compliance engine — state eligibility, suppression lists, calling windows — where missing data means blocked, not maybe. Qualified leads route to partner audit firms by state coverage and capacity, with an audit trail on every decision the system makes.

A generator expands services × cities into a full page matrix with rotating content templates so no two pages read the same, schema.org markup on every page, and chunked sitemaps. Inbound leads — form fills and calls — validate and route to the renting business with a fallback chain. Built on Next.js; the open repo is the generation and routing engine.

An audit engine runs QA checks over an agency's live call transcripts — disclosures, objection handling, dead air, abrupt endings — and every finding ships with a receipt: the exact transcript lines that triggered it. Reports aggregate per call and per campaign. Anything that could touch a production system sits behind a dry-run-by-default action gate.