About ekoreva

I built ekoreva because I was annoyed. Not at a competitor, not at a market gap I'd spotted in a spreadsheet. Annoyed at my own replies on X, and annoyed at every AI reply tool I tried that made the problem worse instead of better.

Here's what was happening. I'd see a tweet worth replying to, open a tab with ChatGPT, paste it in, ask for a reply, and get back something that sounded like a LinkedIn post. Polished, generic, and completely not me. So I'd rewrite it by hand, which took longer than just writing the reply myself in the first place. The tools that were supposed to save time were adding a step.

I tried the reply-generator extensions too. Same problem, just faster. They'd all been trained to produce a kind of default "helpful assistant" voice, and no matter what tone I picked from a dropdown, my replies started sounding like everyone else's replies. Confident, a little bland, missing the specific way I actually talk (short sentences, some sarcasm, a habit of trailing off instead of wrapping things up neatly).

So I started building something for myself first. The idea was simple: instead of asking me to describe my voice in a settings panel, just read what I've actually posted. Not ten tweets. Hundreds. Enough to catch the patterns I don't even notice I have. That became the voice profile at the center of ekoreva, and it's the whole reason the product exists.

What building this taught me

A few things I didn't expect going in.

First, voice is mostly rhythm, not vocabulary. I assumed matching someone's writing style meant matching their favorite words. It's actually much more about sentence length, where they put the joke, whether they end on a question or a flat statement. A model that gets the rhythm right feels like you even when the exact words are different.

Second, context matters more than tone. A reply that's technically "in your voice" can still be wrong if it ignores what's already been said in the thread, or repeats a point someone else made three replies up. That's why ekoreva reads the full thread and the replies already posted before it suggests anything, not just the original tweet in isolation.

Third, people don't want more options, they want the right one faster. Early versions gave five reply suggestions. Almost nobody scrolled past the first three, and half the time people picked whichever one had the highest voice-match percentage without reading the others closely. So that's what shipped: three suggestions, a voice-match score, injected straight into the X compose box.

Still one person

ekoreva isn't a company with a team and a roadmap committee. It's me. I write the code, answer the support emails, decide what gets built next, and use the product every day to reply to my own tweets. There's no outsourced support queue and no product manager translating feedback through three layers before it reaches whoever's shipping the fix.

I think that's a feature, not a caveat. When something breaks, I usually know within hours because I hit it myself first. When someone emails with a weird edge case, I can go look at the actual code that handles it the same afternoon. There's no bigger company here promising a "comprehensive" everything-app. Just a specific tool that solves one problem: sounding like yourself, faster, in the reply box.

Payments run on crypto (USDC across a few chains) partly because it's simpler for a one-person operation to manage than a traditional payment processor, and partly because it fits the kind of people who use X the way I do. The extension is architecturally blocked from reading your DMs, not because of a policy I wrote down somewhere, but because it just never touches that part of the page at all.

If you've read this far, you probably care about the same thing I do: replies that sound like an actual person instead of a chatbot wearing your username. That's what I'm building, one release at a time.