Case studyGenerative AI Engineer
The AI-APPs pipeline — Independent projects (2024–today)
A governed factory to build applications, agents, research lines and dashboards: two houses, one writer per house and written approvals, so that the value also stays in the ability to build the next piece.
- Claude Code
- Next.js
- Zod
- Supabase
- gitleaks
By the numbers
2
houses, with a single writer each
2
written approvals before building each application
4
continuous integration jobs on every change
Context
The portfolio is not built as a succession of independent projects: applications, agents, research lines and dashboards come out of one pipeline, governed by common rules to prioritize, define, build, validate, document and publish each piece.
The challenge
Taking advantage of artificial intelligence without diluting responsibility for the result: nobody being the bridge between the plan and the code anymore, and quality not depending on remembering how the previous piece was built.
How I did it
Two houses, one writer per house
A private planning house keeps the brief, the product vision, the sprints with their plan and their retrospective, the method and the standards; each app's repository keeps the code, the tests and the summary of each sprint. The planning house reads the repositories read-only and never writes in them; synchronization is a read of git, not copy and paste.
Two approvals before the code
No application advances without two written approvals: the current priority, which shows why that piece deserves capacity over others, and the product vision, with its purpose, its users, all its features and its limits. Without them, the pipeline refuses to plan or to open a sprint: conceiving is cheap; building is earned.
A factory agent with human owners
A factory agent keeps the state of the initiatives, verifies the inputs, applies advancement rules and identifies documentation gaps. It cannot approve a priority, modify a purpose, accept a risk or declare a piece finished, and it does not edit the method or the kit without my explicit approval.
Code first, AI when it is justified
Generative artificial intelligence does not come in by default: I must first show what demands interpretation, generation or contextual retrieval, and why a deterministic solution is not enough; activating it requires a written architecture decision. Where there is a model, it is an optional feature, with a deterministic fallback.
Every control is demonstrated by failing
Almost all the rules were born from a red. A secrets sweep that reported "all good" let a weak bait through twice; since then, a gate that was never seen failing does not count. Every change goes through four continuous integration jobs, and every cycle closes with a summary that the planning house reads before planning the next one.
Impact
- Applications, agents, research lines and dashboards built with the same method, at a real cost of US$0 a month.
- Nobody is the bridge anymore: the previous flow cost between 4 and 8 hours of coordination per sprint.
- What one sprint learns in one app travels stamped to the next, in a versioned kit.
- Decisions of purpose, priority, risk and acceptance remain explicitly in human hands.
The goal is for the value not to remain only in the finished product, but also in the ability to build the next one with greater clarity, consistency and speed.