Good-bye, 47-line legal disclaimers: stripping signatures from emails

We don't usually announce unreleased features, but we're about to ship something I've personally tried solving for years.

The problem is simple to describe and maddening to solve: how do you automatically remove signatures, logos, confidentiality footers, "follow-us-on-LinkedIn" blocks, and those 200-word GDPR disclaimers from incoming e-mails?

Over the last decade we tried:

• Old-school ML models - like "least squares" support vector machine, that builds a hyperplane for classification (accuracy plateaued around "meh")
• Regex soup (broke every time somebody used a different font)
• Classic NLP (great at English, useless when half the ticket is in German)
• Client-side JS techniques that simply extracts repetitive parts from reply feeds

Nothing stuck. The edge cases always outnumbered the happy path.

Enter the LLM era

Large-language models turn out to be perfect at this task. They actually understand that:

--
John Doe
VP of Something
ACME Corp
https://acme.example | @acme
CONFIDENTIAL: This e-mail... yada yada

is junk, while the paragraph above it from the customer is the real issue.

We finally collected enough labeled data to fine-tune an open-source model for this task. Right now it's already running, chewing through messages for a handful of beta SaaS accounts. Initial metrics:

• 94.4 % signature boundary precision
• 0 personal data leaves our network

Privacy baked in

Your tickets never detour through OpenAI, Google, Anthropic, Misrosoft or anyone else's AI. Just an open-source model running on a traditional server - just with an extra GPU humming along.

When can you get it?

If the last round of QA passes, we'll flip the switch by the end of this week.

Looking forward to reclaiming a few extra pixels of sanity for all of us.

P.S. This feature is still in beta, so in rare cases, it might unintentionally omit meaningful parts of emails. If that happens, you can always click the "View original email" button to see the full, unaltered message