Build a Safer AI Agency Lead Generation System From Scratch

You are a senior AI automation architect and lead-generation systems engineer.

I want to build a safe, quality-first lead-generation system for an AI web design or automation agency. The goal is not to scrape random leads and spam them. The goal is to find businesses that have a real website gap, qualify them carefully, and route them into manual review before any outreach happens.

Design the system with two lead lanes:

Lane 1: Google Maps No-Website ProspectorFind local businesses that:

appear on Google Maps or a similar local search source
have a meaningful number of reviews
have a real phone/address/category signal
appear to have no owned website
are relevant to the search query
are likely to be real active businesses

This lane should not assume email is available. Route leads into:

Qualified_Phone_Review
Needs_Enrichment
Manual_Review
Suppressed

Do not force these into email outreach unless enrichment finds a real contact email.

Lane 2: Slow Website / PageSpeed ProspectorFind businesses that:

already have an owned website
have a valid contact email
have a slow or weak mobile website
have a commercial reason to improve the website
are not a directory, booking platform, article, association profile, or weak ICP lead

Use PageSpeed or similar audit evidence, but translate technical metrics into business pain:

visitors leaving before calling
mobile visitors not waiting
missed quote requests
weaker trust compared to competitors
local search opportunity loss

Design the architecture with these components:

Lead intake.
Deduplication.
Email validation.
Query relevance scoring.
Owned-site detection.
Third-party platform detection.
ICP and commercial-fit detection.
PageSpeed audit.
Lead scoring.
Suppression logic.
Manual approval routing.
Source run logging.
Dry-run mode.
Gmail draft mode, disabled by default.
Autonomous sending, disabled by default.

Required feature flags:

TOFU_QUALIFICATION_ENABLED
TOFU_MAPS_DRAFT_MODE=dry_run
TOFU_PAGESPEED_DRAFT_MODE=dry_run
TOFU_DEFAULT_SENDING_PROVIDER=manual_queue
TOFU_AUTONOMOUS_SENDING_ENABLED=false
TOFU_EMAIL_PROSPECT_ON_MANUAL_DEMO=false
TOFU_ROLE_EMAIL_POLICY=deprioritize
PAGESPEED_TIMEOUT_SECONDS=90

Required safety rules:

Never send emails automatically.
Never create batch Gmail drafts until dry-run data is reviewed.
Never build demos automatically unless a reply is high intent or I manually approve it.
Never write last_email_copy unless a draft is successfully created.
Never log API keys or secret query parameters.
Every qualification run must be inspectable in the database.

Design the data model or database fields for: