The verdict up front: I didn't take the AWS Gen AI certification.
I'm not lazy — I finished the whole course. The Udemy one, Ultimate AWS Certified Generative AI Developer – Professional, start to finish.
Finishing it made one thing clear: for me, the added value wasn't as big as I'd expected.
The IAM, VPC, and deployment parts — I'd already seen those preparing for the SAA.
The genuinely new parts — Bedrock, guardrails, RAG evaluation — honestly, a weekend with the official docs would probably cover them.
So why spend another few weeks grinding a question bank and booking an exam slot?
I didn't. I put that time into building products instead.
The real learning started after I began building.
What only building teaches you
I built a few AI products. Two of them are SpeakUp and AI Video Cut.
On the surface they're the same kind of thing — both LLM-driven, both handling AI input and output.
But once you actually build them, they're worlds apart.
SpeakUp is real-time voice streaming: it needs streaming, low latency, and to return a score instantly. It does speaking assessment, so one wrong call directly damages user trust — the whole pipeline revolves around error rate.
AI Video Cut turns a prompt into video: it's an async job that's allowed to be slow. The point isn't speed — it's getting closer, round after round, to what the user wants. It can tolerate error, because that's inherently an iterative process.
One breaks trust the moment it's wrong; the other only gets good by being wrong and fixed, again and again.
Just from that, the two products' architectures — and even what "good" means — are completely different.
A cert hands you a "standard architecture diagram for LLM apps." But build for real, and you find there's no standard architecture.
That kind of judgment isn't in the question bank.
This field moves faster than a cert can
There's a practical reason too.
When I'm building, something new drops every two or three days — a new model, a new API, a new approach.
A cert validates a snapshot from a few months ago. Products push you to chase what's moving every week.
Those two rhythms are just too far apart.
Half of it is old, half is too new
I eventually worked it out: the hard part of an AI product sits right on both sides of the cert.
Half is traditional system architecture — DB, queue, cache, the whole backend. I learned that preparing for the SAA. Not new.
The other half is the LLM, changing every week, faster than a cert snapshot can follow.
One half old enough that I already know it, the other new enough that the cert can't keep up. The cert lands squarely in the middle, where I need it least.
So I chose to build
So I didn't take the exam.
Not because it has no value — I learned what there was to learn, no loss there.
And a cert and building aren't either/or anyway. I just only had time for one thing back then.
I chose to build.
Because what proves I can build AI products was never a badge. It's the things I actually shipped — like SpeakUp.