Small models for small jobs

(or, a funny thing happened on the way to the data center)

When the Tailor beta launches, the app will use a small model from Arcee AI called Trinity-Mini-Base, a 26B model with a mixture-of-experts layer that helps it work efficiently even at its small size.

Why Trinity-Mini?

• Small. Runs affordably at scale.
• Fast. Low latency even with cold starts.
• Stable. This matters when you’re figuring out what’s broken in your app.
• Open source. The Trinity family has an Apache 2.0 license and open weights.
• Training data provenance. Trinity was trained in partnership with Datology.

Anyway, I encourage you to read The Trinity Manifesto.

Core AI philosophy: Small models for small jobs

You don’t need a 600B model to break down your task list. You don’t need to burn down the rainforest to unpack your sprint.

A small model, trained well for its specific job, can nearly match a large generalist in domain expertise. I proved this with Lenina.

Lenina is my fine-tune of Trinity-Nano-Base (6B), trained on approximately 31,000 data pairs to be a publishing assistant. It learned US copyright law. How to fill a form TX. How to apply for CiP data. How to choose BISAC codes. How to create structured metadata. And it hit 68.5% semantic similarity on complex tasks when I validated it against a held-out set of 1,543 data pairs. To put that in perspective, it performed within 2.5 percentage points of a 399B model even though it has 1/66th the parameters.

The 399B model runs in a data center. Lenina runs on my desktop PC.

Why does that matter for Tailor?

• Inference costs. Because small models are cheaper to run, I can keep Tailor free with few, strategically placed ads (so you never see an ad when you’re head down working—productivity poison). And the paid pro version, which unlocks additional features and removes ads, stays affordable.
• Carbon footprint. I don’t want to boil the ocean reminding you to do your expense report. Smaller models mean a smaller environmental impact.
• Data security. Tailor runs on RunPod’s infrastructure. We never hit an inference provider’s API. Your data stays where it belongs—with you.

This isn’t just about efficiency. It’s about empowering the app with AI that I don’t lose sleep over.

The tailor-made version for the app

What’s going to make Tailor special is the model I’m training right now.

The Tailor model I’m training will know ADHD, not as a label, but as a lived reality. It will have behavioral training specifically designed to:

• Support without controlling.
• Encourage without condescending.
• Break down tasks without overwhelming.
• Surface trends from your focus data (“You focus best Tuesday mornings”).
• Make helpful, contextual suggestions (“It’s Friday afternoon. That’s historically your best time for admin work”).

Not a generic productivity bot. A focus companion that understands how a brain with ADHD works.

The productivity apps I’ve tried that market themselves as “ADHD friendly,” frankly, are not. They’re generic productivity apps. The assumption is that if it’s a productivity app then it must be for people with ADHD (because we don’t focus well). That’s not what Tailor is. Tailor—the app and the model—is built from the ground up to meet the challenges that adults with ADHD really have.

Why I’m the right person for this

I’m not an ML engineer. Or any kind of engineer. I’m a writer, editor, and publishing professional who taught myself fine-tuning because existing models, however big they get, do a handful of things that make them utterly impossible for me to use.

• They talk to me like I’m stupid (looking at you, ChatGPT).
• They tell me, “You’ve worked hard today, go take a well-earned rest.” (Please never tell me to go rest when I’m motivated to focus.)

(Oh, and rest is not “earned.” Rest is a human right. But go on.)

• They personify themselves or impersonate humans. Ever had one say, “I get frustrated when that happens to me, too.” No you do not you are a highly sophisticated autocomplete algorithm not a person.
• They’re mostly accessible through chatbot wrappers that push them to engage users. Those bait questions at the end of every response? Designed to keep you engaging with them instead of doing what you need to do.

So, like I said, not an ML engineer. But I’ve done this before. Lenina wasn’t just a proof of concept. It was an evaluated, documented success. I know how to:

• Curate training data across behavioral categories.
• Hand-write examples to establish voice and values.
• Balance topical representation to avoid repetition.
• Teach a model epistemic humility (that is, to “I don’t know” when it doesn’t know).
• Evaluate training success against untuned baselines.

Tailor is the same bet, just in a different domain.

What’s next

Trinity-Mini-Base will take Tailor to beta. The custom model, when it’s ready, will take Tailor to the next level. I expect that when beta begins, you’ll see a productivity app, a focus companion, that works well and has some novel features (the intention → sprint → unpack core loop). But when we make the switch to the custom Tailor model, you’ll see the app come to life (in a good way, not in a creepy Frankenstein way).

I’m working on the training data now. I’ll share progress as I go: what’s working, what’s failing, what’s taking longer than expected, and so on.

If you’ve ever wished a productivity tool actually understood how your brain works, that’s what I’m working on.

Beta signups are open.

—Catherine