Doorman: my apps' contact form got 23 messages in six weeks, and all 23 were pitches

Description and use

Doorman is a browser‑based tool that sorts contact‑form emails from Shopify app listings, keeping genuine customer inquiries on top while folding away sales pitches, using EmbeddingGemma and logistic regression for classification.

Problem

Contact form messages from Shopify app listings were all sales pitches, burying genuine customer inquiries.

Outcome

In a six‑week test, Doorman folded away 102 of 130 pitches while only 2 of 115 customer messages were mistakenly hidden, achieving high precision for real inquiries.

Workflow

• Load messages by pasting, dropping .eml files, or uploading a Gmail .mbox
• Press "Sort Inbox"
• EmbeddingGemma 300M processes each message and its sentences into 768‑dimensional vectors
• Logistic regression classifies each sentence with a pitch probability
• Messages with probability ≥0.8 are folded away, 0.5–0.8 marked "not sure", below 0.5 remain as customer messages
• Highlighted giveaway sentence is shown for folded pitches

Evidence

Between 29 August and 13 October their listings' contact form and my email got 35 messages: 23 once the repeats and image‑only ones are gone, and all 23 were selling something.

Source

Doorman sorts a contact‑form inbox: customers first, pitches folded away.

Source

EmbeddingGemma 300M, running in a Web Worker on your GPU, turns each message and each of its sentences into 768 numbers; a logistic regression trained on 245 hand‑labelled forum posts and my…

Source

Five‑fold cross‑validation over the forum posts, split by topic so nothing is ever scored by a model that saw its thread (full report).

Source