L.U.C.I.A.: al detalle técnico

The article is a written version of my GDG Madrid presentation, considerably longer. The code snippets illustrate our implementation method and design decisions, not a direct copy from the repository. The aim is for you to be able to reconstruct the framework on your own, not just copy ours. On Monday morning, your director read in Forbes that AI reduces development time by 70%.

It's Monday, it's nine o'clock, and they want it by Friday. This scene has repeated in many companies with a recognizable shape: the expectation curve grows exponentially while real engineering capacity grows linearly. The gap between these two curves closes through one of two ways: either you build a framework to withstand the pressure, or the team pays the difference in technical debt, incidents, and burnt-out personnel.

We chose the former, and we named the framework L.U.C.I.A. This is what's inside. The diagnosis: what exactly happens when there's no framework? It's not that AI doesn't function. It functions too well for how quickly it's adopted. There are five real failures we've experienced, each followed by a countermeasure: 1. Security and PII leaks.

Personal API keys in local scripts and client data sent without anonymization to public clouds, because someone pasted an entire CSV into the prompt to help it understand better. 2. Cost inflation. Without observability of expenses, there's no way to explain why the bill has tripled. It's usually the same: huge contexts resentfully forwarded in their entirety on each turn and redundant requests not deduplicated.

3. Technical degradation. The model optimizes for compilation, not for maintainability within six months. Spaghetti code, logical hallucinations, and debt that accumulates faster than it's paid off. 4. Architectural fragmentation. Each generation ignores previous project decisions. Clean Architecture and SOLID evaporate because the agent doesn't know they exist.

5. Shadow AI. Each person has their own model, prompt, and criteria. There's no corporate brain, no repeatable flows, and knowledge doesn't accumulate anywhere. The pattern: none of the five is a problem with the model. They're process problems. And process problems are solved with processes. Chaos is the enemy of scale. What's missing isn't more AI: it's a control framework.

The five L.U.C.I.A. decisions are not a tool you install. They're an agreement on how the team works: Lifecycle — covering the entire project lifecycle from idea to deployment, ensuring all roles work orderly with AI. Not just development. Universal — requesting a logical model, not a provider: a single output, a single policy, and adapting to the CI/CD tools you already have.

Collaborative — people and agents on the same board: AI proposes, people decide. Iterative — small tasks, branch and merge request per task, verification in beta before production, continuous improvement. Automation — everything repeatable is a command; deterministic things never delegate to AI. It's DevSecOps culture, not magic.

If you only take one sentence from the entire article, let it be this: AI proposes and prepares; from Merge Request onward, everything is deterministic. That sentence is a boundary, and placing it well is 80% of the work. As for a single guiding principle, remember this: AI accelerates the cognitive part, understands, plans, writes. The rest, from build, tests, image, deployment, promotion, is entirely deterministic.