Case Study · AI engineering methodology

The AI translates and executes. Decisions and verification stay human.

Cadensa is a live SaaS product since 2026-05-31 with 180k+ lines of code, built with AI agents. Every brand and business decision, and every production-measured check, stays in human hands.

ProductCadensa, live SaaS since 2026-05-31
Codebase180k+ lines, 6 repos
Automated tests187+
Security vulnerabilities fixed96 (across 6 repos)

The methodological question

How do you trust an AI agent without trusting it blindly?

The AI translates a spec into code correctly, but the spec itself can be wrong, and correct code doesn't mean correct behavior. Only measuring settles that, not reading.

The development's real output

Not theory, but dated, verifiable work on a live product.

187+ automated tests (backend + frontend)
1,945 npm packages audited, 0 critical license conflicts
96 security vulnerabilities fixed across 6 repos, including a CORS bug validated in production
11 active + 22 completed development roadmaps
Every claim verified in production, not from reading code

Where human oversight mattered

Four moments where plain trust would have failed.

  1. 01

    The spec itself can be wrong

    The AI computed against the doc's assumed brand color, but Cadensa's real brand color was different, so it had to re-measure live.

  2. 02

    Technically correct isn't always the right call

    The AI suggested a darker brand color to fix a contrast bug. The human chose to keep the brand color, and the AI found an equally valid alternative fix.

  3. 03

    Code can be correct while behavior isn't

    A reduced-motion fix was syntactically correct, but measuring it live showed only the animation was neutralized, not the underlying value. Only a real browser check caught it.

  4. 04

    Nothing is done until it's measured live

    No “done” report shipped without an actual browser check. Code that “looks right” isn’t enough.

What follows from this

AI speeds up execution. It doesn't replace responsibility.

human sign-off on every decisionmeasured, not assumed, outcomesa documented decision log

Next step

Building with AI agents, but want control?

Cadensa's development proves exactly this: AI speeds up execution, while every decision and verification stays in human hands.

Let's talk