Write what you meant.

Delv turns a fragment into the message you would have written. To this person. Not to everyone.

Read the research behind how we tell the difference.

Join waitlist See why ↓
What you typed
tell him no but keep the door open|
Who it's going to

Delv knows

Fluency is free now.

Which is exactly why it stopped being worth anything. Every draft arrives well-formed, and readers have learned what well-formed and empty looks like. Sounding like a machine now costs you something measurable: replies you don't get, reviewers who stop reading, people who can tell you didn't write it.

✓The tell was never bad grammar. It's fluent and it's empty.
✓Machine text is written for a general reader, so it explains what the two of you already know.
✓What's scarce now is the thing that can't be averaged: your voice.

Delv knows you

How you write

  • Learns from what you've already sent, not from a prompt
  • Your rhythm, your length, your openings, your habits
  • Including the parts you'd never think to describe
Join waitlist →
You, 9:04amtwo words, no greeting
You, draftingyou never say "circling back"
You, annoyedsentences get shorter
You, asking a favoryou get funnier, not softer

Delv knows them

And who you're writing to

  • You don't write to your lawyer the way you write to your cofounder
  • What you both already know stays unsaid, the way you'd leave it
  • Everyone else models the writer. We model the pair.
Join waitlist →
Youclipped, no preamble
→ cofoundershared context assumed
→ lawyercomplete sentences, dates
→ best friendbarely punctuated

How we got here

Built by someone who watched it happen.

Delv started as a research tool. Then the entire category got commoditized overnight. What I learned building through that is what Delv is now.

2021

A research tool, launched at 15

I was a research intern doing EEG analysis, spending most of my time just finding the right passages in a massive literature. So I built a synthesis engine on GPT-3 that could cut literature review time by 75%. Launched on Product Hunt on my fifteenth birthday, hit #3 product of the day. 20,000 people joined the waitlist.

2022

Then ChatGPT shipped

The premise of the company was that summarizing large text collections was hard and scarce. Within months, general-purpose LLMs made it neither. I watched the thing I built get commoditized from the inside.

2025

Built an AI email client at YC

Cofounded Slashy, an AI-native email client (YC S25). Working on machine-drafted email in production taught me the real failure mode: the drafts didn't fail because they were badly written. They failed because they didn't sound like the person sending them.

Now

The same problem, from both ends

Both experiences pointed at the same thing. Once machine text became fluent and everywhere, fluency stopped being worth anything. What became scarce is voice. So Delv models the pair: how you write, and who you're writing to.

Delv beyond

We're publishing the research.

There is currently no accepted way to measure whether a piece of text reads as a specific person writing to a specific person. Existing detectors work at the level of the document and don't attempt it. We're building that measurement in the open, because we need it before we can claim anything.

In progress

A taxonomy of machine register

The measurable features that make writing read as machine-generated, at the lexical, syntactic, rhetorical, and pragmatic level.

Next

An open benchmark for voice fidelity

A public dataset and evaluation harness for writer-specific and recipient-specific fidelity. Usable by anyone, including people building against us.

Next

The blind recipient study

The only test that matters: can the actual recipient tell? Not a classifier. The person you sent it to.

Say it the way you would have said it.

"It's not that it wrote badly. It's that it wasn't me."

We'll let you know when Delv is ready. No spam.

You're on the list. We'll be in touch.

Get in touch

Contact

pranjali@delv.ai