The problem: school administration with no digitization

The Centro Educativo en Arte Benito Juárez ran all of its administrative processes in a fragmented way. As a multimedia teacher at the institution, I experienced that friction firsthand, which completely changed how I designed the solution.

Screenshots of the system in production

System architecture

Axula uses a modular architecture of 20 Flask Blueprints organized by functional domain. SQLite in WAL (Write-Ahead Logging) mode allows concurrent reads without locking while the system writes — key in a school environment with many simultaneous users.

Main modules

Tech stack

Backend

AI and document generation

Frontend and security

How Axula was built

I built Axula from inside the institution — as a teacher living the problems I was solving. That completely changed the design decisions.

1. 1
   Direct observation (2 weeks) Before writing code, I documented every manual process: how report cards were filled out, how long attendance took, where information got lost between teachers and the secretary's office.
2. 2
   Domain-based Blueprint architecture I designed each module as an independent Blueprint. This lets the system grow without touching what already works — a new module doesn't break existing ones.
3. 3
   Iterative development with real users The attendance interface was redesigned 3 times until any teacher could complete it in under 90 seconds per classroom. Not an estimated speed — time measured with a stopwatch.
4. 4
   AI integrated into existing flows, not as a separate feature The LLaMA 3 assistant appears at the exact moment a parent needs an answer — not as a generic floating chatbot, but built into the parent portal with school context.
5. 5
   Full modularization (20 Blueprints) Refactoring from a monolith to a Blueprint architecture was the most critical change. It lets different team roles work on modules without code conflicts.

What I learned building Axula

Building a system you use yourself changes how you design. You're not interpreting a brief — you're solving your own problem. Every UX decision has real consequences: if the attendance form is slow, you lose the time. That forces honesty.

The most important lesson: AI isn't a feature, it's a layer. The LLaMA 3 modules work better when integrated into flows that already exist than when presented as separate tools. The query assistant works because it appears at the exact moment a parent needs it, with real school context — not as a generic chatbot.

The second lesson: SQLite in WAL mode solves 90% of concurrency cases without the complexity of Postgres. For internal systems with dozens of simultaneous users, it's the right choice. Adding unnecessary complexity is the most common architecture mistake.