Nuzic: LoRA and fine-tuning for an educational musical AI

Music through numbers

Nuzic is an educational project born in Barcelona that sets out to understand, create and feel music through mathematics. Its own system describes rhythm and sound with numbers, and it comes with a free, open-source online app, a classroom program (PlayNuzic), an album composed entirely with the system and a documentary, ¿Cómo suenan los números? (“What do numbers sound like?”).

I’ve been part of the tech team since 2022, looking after the website, the database behind the music app and all the DevOps. In 2026 the question changed: if no AI model knows the Nuzic system, can we teach it?

A LoRA that speaks Nuzic

No model on the market can explain Nuzic’s mathematics-based musical system. It lacks the concepts, the vocabulary and the way of reasoning. That’s why we’re doing fine-tuning: a LoRA trained locally on Nuzic’s knowledge, so the AI can explain concepts, solve exercises and answer the way one of the project’s teachers would.

Why LoRA. Retraining an entire model makes no sense. A LoRA is a lightweight adapter that’s trained on your own hardware, versioned like any other file and can be combined with RAG to cite the original theory.
The full pipeline. A curated dataset, fine-tuning with Unsloth or Axolotl, merging, quantization, deployment on the lab’s model gateway and evaluation with real questions from teachers.
Truly educational. The model tells the Nuzic system apart from classical music theory and adapts its tone to a student, a teacher or someone discovering it for the first time.

The dataset factory

Good fine-tuning starts with the data, not the GPU.

Ingestion of the theory PDFs and the websites’ content, converted to JSONL.
Synthetic generation of pairs with explicit reasoning, in five types: calculation and conceptual reasoning, contrast with classical theory, direct answers, multi-turn conversations and answers with citations.
Three tones for the same content (technical, classroom and general-audience), with a stable identifier per question.
Automated quality control of duplicates, length and reasoning consistency, with cost metrics per batch.
Cost optimization: switching from attaching the PDF to every call to using prompt caching made it about nine times cheaper.

Before training, a person reviews every pair on a curation website with roles (administrator, creator and validator), editorial statuses, auditing and export ready for Unsloth and Axolotl.

Websites that run on AI

The AI doesn’t stay in the lab. Parts of Nuzic’s websites are already built with it and run on it.

Migrating nuzic.org to Next.js, done by Pi agents: an orchestrator and subagents that each work on their own files, with a looped visual review against production (±2 pixels).
Educational content prepared with agents and reviewed by the team before it’s published.
A private assistant with Nuzic’s context for the team, which looks up the theory with citations and doesn’t send any data to third parties.

A virtual company of agents

On Discord, Hermes Nuzic works as a virtual company with a product lead, an architect, backend, frontend, technical and visual QA, DevOps and a release guardian.

It only works on allowed branches.
It runs tests and an automated bug review before every merge request.
It only merges with human approval verified through the GitLab API.
It looks up the theory in an Obsidian vault, with citations.

What I learned

An in-house model starts with well-curated data. The GPU comes last.
With a LoRA, a small educational organization can have its own specialized AI, without depending on anyone and with its data kept in-house.
Development agents work when they have tightly scoped permissions and a clear rule not to touch production.