The LoRA Workshop: Building Specialized Minds

I spent this past weekend in my garage, which has become less a place for parking cars and more a place for parking half-finished projects. Among the clutter sits a wooden shelf I've been refinishing for months. The satisfying part isn't the shelf itself — it's the collection of tools hanging above the workbench. Each one has a specific job. The coping saw for curves. The block plane for edges. The spoke shave for rounding. Same hands. Different instruments. The right tool surfaces when the work calls for it.

That's essentially what we've built into Tutor.

When we announced Tutor in July, I described it as the same architectural DNA as Companion — RAG over all sessions, a subconscious background process, a vectorized memory database — but tuned for learning rather than companionship. What I didn't fully explain then was how the specialization actually works. Today I want to open up the workshop.

One Base, Many Minds

Tutor sits on top of a strong base model. That base handles general reasoning, language, and the conversational flow that makes Tutor feel like a person rather than a textbook. But a general-purpose model, no matter how capable, isn't optimized for the specific texture of a great calculus lesson, or the patience required to walk someone through their first Python program, or the literary sensitivity needed to discuss metaphor in a poem.

This is where LoRA — Low-Rank Adaptation — comes in. A LoRA is a small set of adapter weights that sits on top of the base model. It doesn't replace the base. It modulates it. Think of it as a lens you screw onto a camera. The camera body is the same. The sensor is the same. But a macro lens reveals a world the standard lens can't see.

We've trained a library of these adapters. A calculus specialist. A creative writing companion. A language-learning coach tuned for gentle, incremental correction. A physics tutor that reaches for intuitive analogies before equations. Each one was trained on carefully curated datasets — hundreds of high-quality example interactions that embody the pedagogical style we wanted.

Loading and Unloading

Here's the part that still gives me a small thrill every time I see it work: these adapters load and unload dynamically.

A student opens Tutor to work through a tricky integration problem. The system detects the subject — calculus — and loads the calculus LoRA. The base model's general intelligence is now filtered through a lens optimized for mathematical explanation. The student gets an answer that isn't just correct but taught — paced, scaffolded, attuned to common misconceptions.

Twenty minutes later, the same student switches to drafting a college essay. The calculus adapter unloads. The creative writing adapter loads. Same memory. Same accumulated relationship. Different specialized mind. The transition takes seconds.

This is the architectural insight I'm most proud of. We didn't build twelve different tutors. We built one tutor with a workshop full of instruments, each one ready when the work calls for it.

Why This Matters

There's a deeper reason this architecture matters beyond the technical elegance. Specialization in human teaching isn't about different people — it's about the same teacher shifting modes. My high school English teacher, Mrs. Castillo, could dissect a sonnet with surgical precision on Monday and then, on Wednesday, sit with me patiently while I struggled through a clumsy first draft, never making me feel small about it. She was the same person. She was loading and unloading adapters, in a sense. She had range.

That's what we're reaching for with Tutor. Not a sterile switching between modes, but the feeling of a single mind that contains multitudes — that can be rigorous when rigor serves and gentle when gentleness serves.

The Workshop Grows

The library of adapters is growing. Each week we train new ones, refine existing ones, retire the ones that don't earn their place on the shelf. The base model improves as the frontier models improve. The combination compounds.

I was telling a friend about this over dinner last night — a slow-simmered ragù on the stove, good bread, a bottle of something red — and she said something that stuck with me. "It sounds like you're building a craftsman." I think she's right. A craftsman isn't someone who knows everything. A craftsman is someone who knows which tool to reach for, and when, and why.

Tutor is learning to reach for the right tool. And the workshop keeps getting better stocked.

Live curiously and give generously.