Demo — AI intake assistant (email triage + drafted replies)

Demonstration project — every email, name, and address here is synthetic. The company ("Maple Ridge Plumbing") is invented. What's real is the pipeline and the outputs you can inspect in this folder.

The problem this solves

Service businesses run their whole front door through one inbox: quote requests, "where's my invoice", emergencies, reschedules, job applicants, and vendor spam all land in the same place. Someone triages it between jobs, urgent things wait behind spam, and every reply is typed from scratch.

What the assistant does

For each incoming email, the pipeline:

Classifies it — quote request / urgent service / billing / scheduling / hiring / follow-up / vendor-or-spam, with a priority attached (urgent = same-day).
Extracts what a dispatcher needs — address, phone, subject of the work — into a triage log (CSV here; the client's CRM or sheet in production).
Drafts the reply in the business's voice, ready for a human to approve. Vendor pitches and SEO spam get filtered, not answered.

Nothing auto-sends. Drafts land in a review queue; a person approves, edits, or discards. That human gate is deliberate — it's what makes the assistant safe to put in front of real customers, and it still removes ~90% of the typing.

What's in this folder

make_inbox.py — generates the 10 synthetic emails (so the demo is reproducible).
intake_assistant.py — the pipeline. In production the classify/draft steps call an LLM with a business-specific prompt on the client's own API key; in this public demo those steps are a deterministic stand-in with the identical interface, so anyone can run it end-to-end without a key.
inbox/ — the 10 raw emails.
triage_log.csv — the machine-readable triage output.
replies/ — the 8 drafted replies (spam gets none).
demo.html — the human view: the triaged inbox with chips, extracted fields, and expandable drafts.

The honest part

AI does the reading, sorting, extracting, and drafting; a human approves every message that goes out. We disclose that in every deployment — it's why setup is days, not weeks, and why the price is fixed. A deployment is tuned to your categories, your voice, and your tools (shared inbox, CRM, job tracker), and comes with a runbook plus 30 days of prompt fixes.