NowSDET and AI automation at a payments fintech
I'm Andres Carreon, a senior SDET who writes a lot of code. For nine years I've tested iOS, Android, desktop and web apps so they ship without surprises. Lately I build AI agents and the evals that keep them honest. After hours I build my own tools and write down what I find.
A recorded jevvium run. Goal: A registered user logs in with a valid email and password and is told they are logged in. Four decisions: tap Login, type the email and password, tap LOGIN, goal reached. Passed in 5.2 seconds, and the replay with plain Appium passed too.
// generated by jevvium with jev-1.13.0
describe('login', () => { it('A registered user logs in…', async () => { await $('~Login').click() await $('~input-email').setValue('qa.demo@example.com') await $('~input-password').setValue('Str0ngPassw0rd') await $('~button-LOGIN').click() await expect($('-ios predicate string:label == "You are logged in!"')) .toBeDisplayed() }) })
Fixed selectors, real assertions, no AI at run time. It runs in CI like any other test.
Now building · open source
You describe what a feature should do. A decision model finds the path through the app one tap at a time, and jevvium writes it down as a plain WebdriverIO test that CI runs with no AI in the loop.
Read the code on GitHub1Describe it
A goal, some test data and what should be on screen when it works. A few lines of YAML.
2Explore once
A decision model picks the next action from elements that really exist on screen, and stops to escalate when it isn't sure.
3Run forever
The path becomes a normal test. jevvium replays it once with plain Appium to prove it passes, then CI owns it.
47.7s → 16.3s
to explore four criteria, after rebuilding how it reads, decides and taps
~25 ms
per simulator tap through a small native helper, instead of about 400 ms
< $0.001
of model cost for those four criteria
4 at once
simulators exploring in parallel, one Appium server each
Field note: it found a real bug
In Sauce Labs' demo shop, “price, low to high” puts $7.99 after $49.99.
The app compares prices as text. The model couldn't see prices, so it wasn't sure the goal was reached. The expect check is what caught it.
Day job
Quality engineering at a global consultancy, a crypto wallet used by millions, and now a fintech startup. Here it is as release notes.
v3currentAug 2026 to now
Software engineer, quality and AI automation
name under NDA
v2May 2022 to Aug 2026
SDET, quality engineering
self-custody, used by millions
v12017 to May 2022
QA engineer, test automation
enterprise systems
After hours
Nine shipped so far, two with paying customers. Building them is where most of my AI tooling comes from, including the agent that triages my own bugs.
Plus two open-source test frameworks, a hackathon soap opera and a few experiments.
All projectsWriting
Things I found interesting enough to write down: AI, attention, and building things alone. Most of them in English and Spanish.
Jev is a text-only decision model, so I built it a retina out of plain code: letter maps, legends and silhouettes drawn with # and dots. Four days later, TypeSafe announced it could see all along.
A builder's journal from the Qwen Cloud hackathon: building Qwen City, a 24/7 town of generative agents whose memories the audience can rewrite — and what the usage dashboard taught me about my own architecture.
More than 80% of companies now use AI, and around 95% of them cannot point to a single dollar of return. The tool is not the problem. Here is why AI dies inside companies, and why that same tool is the cleanest edge a solo builder has ever had.
ENES
Phones train brains to bail at minute six. Almost nothing worth learning fits in six minutes. The gap between those two facts is the closest thing to a free moat available right now.
ENES
A concert made it obvious. Attention became a currency with an exchange rate and tiers of access most people can't see. Here's how the mechanism works and how to use the information.
ENES
