Form Pipeline — AI-Powered Spam Detection & Routing

A multi-layered form submission processing system that screens incoming submissions for spam and routes valid ones to configured destinations. Running in production across multiple sites.

The Engineering Insight

Most AI spam filters send every submission to an LLM. That’s expensive and slow. The architecture here uses AI as the last resort, not the first — and that design decision is the entire reason it works economically.

The Filtering Pipeline

1. ASN/IP-level blocking with decay — tracks spam sources by IP, escalates to ASN-level blocking for persistent sources. Blocks decay over time so legitimate senders aren’t permanently banned. Whitelisting available per IP or ASN.
2. Pre-AI heuristics — gibberish detection, format validation, and behavioral signals catch obvious spam before it reaches the expensive layer.
3. AI content analysis — submissions that pass the earlier layers get reviewed by an LLM (Anthropic or OpenAI) for final determination.
4. Every layer is independently configurable — customers can disable AI and run IP-only, disable IP blocking and run AI-only, or turn everything off and use it as a pass-through router.

Results

1,074 spam submissions rejected, 50 valid submissions routed, total AI API cost of $5.27 over 3.5 months. The cost is low because most spam never reaches the AI layer.

Routing Engine

Once a submission passes filtering, configurable connectors route it to Slack, email, webhooks, ConvertKit, Zoho, or other integrations. Each form can have its own connector configuration.

Cross-Product Integration

Exposes an API that other applications call to check IPs and emails against the flagged database. TC Track uses this at signup for fraud prevention — flagged users get shadowbanned (registration flow looks identical but no account is created).