Today is June 12. I hoped to be in closed, private beta by June 15.
That’s looking unlikely now.
Why? A few setbacks.
First, I had to change my user authentication database. The first provider I tried wasn’t working out, so I rebuilt the entire user database in a new environment. That added days of backend work.
Second—and this one stung—I had to switch the model I’m using for inference.
I really wanted to use Trinity-Mini. I believe in it on principle. It’s a good model. But it’s just not tuned for the kind of chat and instruction-following that Tailor requires. That’s not Trinity’s fault. It’s just not the right tool for this job.
So I switched to Mistral-Small-3.2-24B-Instruct-2506.
Why this model? Two reasons:
- It follows instructions better—critical for Tailor’s use case.
- It speaks more naturally.
What’s the same? The small size. Low carbon footprint. Open source. That hasn’t changed. Right now, for closed beta, we’re running inference over Openrouter. For open beta, Mistral-Small will be hosted in a secure Runpod so user data never hits Openrouter’s API. And training a custom Tailor from Mistral-Small is still the plan. The final product will launch with a fully custom, fine-tuned model developed especially for Tailor’s use case.
What does this mean for beta?
I still anticipate shipping the closed beta before the end of June—hopefully in another week.
I’ve added a lot of features since my last update, and I’m really pleased with them:
- The custom double creator is built and working. I literally love it.
- All the doubles have their personalities fully fleshed out and are usable.
- Three new working environments—rain, campfire, and cafe—in addition to the default silence.
- Three new themes—golden hour, midnight, and riot—to reskin the app to your taste (alongside the original cream-and-teal Tailor theme).
- Tailor can now help with your backlog—tidying it by adding subtasks, and helping you choose what to work on if you’re not sure.
Still in the works:
- Full voice support (STT and TTS), on-device and with Kokoro voice.
- AI-supported insights analysis.
The core loop works.
I could launch closed beta tomorrow. But I don’t want anyone using this until it’s something I absolutely love to use. And it’s not quite there yet.
To paraphrase the immortal words of Shigeru Miyamoto:
A rushed product eventually comes out, but a bad product is bad forever.
You only have one chance to make a first impression. I intend to make a great one.
—Catherine
P.S. If you’re interested in following along, I’ll post another update when the closed beta is ready. No mailing list yet—just watch this space.