Speed to Lead

The end-to-end budget is sub-5 seconds from form submission to first outbound message. Form posts hit a Next.js route handler that immediately enqueues the lead and returns 200 — no blocking work on the inbound request. A Fluid Compute worker pulls the event, runs a Claude call for intent classification and message generation, and fans out to Twilio (SMS) and Resend (email) in parallel. Live chat is pushed over a Supabase Realtime channel for the agent's inbox. Lead scoring is a small rubric-based Claude call (not ML): it scores 0–100 on fit, urgency, and deal size, and returns a routing key. The dispatcher then looks up which team member owns that key and attaches the lead. Everything — submissions, model calls, dispatches, replies — writes to a Supabase event log that powers the real-time analytics dashboard. Key decisions: inbound and outbound are fully decoupled so a slow model call can't drop a form submission; Claude is only used where judgement matters (intent, tone, routing) — templates do the rest; response-time SLA is measured end-to-end, not per-stage, so regressions surface immediately.