I learn by building real things for real people.

Why

I wanted to find a niche where AI would solve something for real. I asked my dad which process took him the longest and repeated the most. He said entering electronic invoices (CFDI) for the monthly tax return.

He used a program that downloads the invoices from the SAT into Excel, but it doesn't classify anything. Every month he built the formulas and the classification for each client by hand.

The hard part wasn't the code

It was understanding the tax logic: how each item is classified and why. At first there were errors I couldn't see because I didn't know enough. Through the project I learned a good part of my dad's job, and every rule in the system had to be backed by the law. In tax work, accuracy comes first, because one mistake hurts everyone.

Since I depended on his knowledge, I also had to work around his schedule.

Trust before features

Most of the work went into making the data reliable and making sure clients' data never leaks. Credentials are encrypted and the dashboard only lets my dad and me in. There are tests that kill the process halfway through to check that nothing breaks, and backups with tested restores.

One story sums it up: sometimes the SAT accepts the request and delivers zero invoices without flagging any error. One client showed $0 in months that did have activity. I built a guard that detects that suspicious zero and stops it before it reaches a tax return.

v1

On October 30 we reviewed two clients together against how they had actually been filed, and the results were quite accurate. The system also surfaces information that used to take digging to find. Now we're checking it against all of October before it replaces the old process.

Film first, AI second

I've always been drawn to traditional filmmaking: how it's shot, the creative process, the idea of directing. AI is what let me get close to it. Without it I couldn't have made this short.

My brother and I had just finished our biggest order, the method was fresh and we had money for credits. I already had the story: a man in a modern war who somehow makes it home. I wrote it listening to Max Richter, whom I used as an emotional guide for the original score.

Letting go of control

I brought over the method from the previous order: Blender video to guide every movement. The first scene, the trench, came out ugly. I realized the model is good enough that I don't need to control every detail: very clear instructions with some freedom. The trench, redone without previz, turned out great.

Directing choices

• The horror is told without showing it. There's no gore at any point.
• The ending changed after it was published. John asked his dog for forgiveness at the grave, and it came across as if he cared more about the dog than about his wife. I tested it with my test audience, my parents, and replaced it with a shot of the two silhouettes against the light.
• To save credits, I tested at 480p and only re-ran approved shots at 1080p.

I was told not to build it

Before building, I ran the idea past a council of five AI critics. Their verdict was to rework it: no app, just set up matches by hand over WhatsApp. I kept their best advice: measure how many pairs of strangers play again without me stepping in. But I saw it was simple to build, and within a day it worked.

A lab before touching the ranking

Before changing the skill-rating system, I simulated three alternatives over more than two million synthetic matches. The one that predicted best was the easiest to game: if you only report the scores that help you, you climb without playing any better.

A tool for the people who decide

I figured the app would grow if coaches approved it and used it with their students. So I built them a dashboard: groups, practices and notes for each player. Mid-practice, the coach taps a student, dictates a note and decides whether to send it or save it. Dictation runs locally, with no paid API, and the audio is deleted.

The learning curve

It was the first time I paid for a Claude plan. I built the first version in Claude Design and made beginner mistakes, like iterating in old chats with too much context. When I exported it, everything loaded from a single, extremely heavy HTML file. Armando told me it was really slow and people would leave. He was right: with Claude Code I split it into files and slimmed it down.

The WhatsApp bot

The bot asks for the date, group size and extras in English or Spanish, saves the lead and hands the conversation over to Armando. He sees everything live from his phone, and if he types, the bot goes quiet. With one command he confirms the booking, and a reminder for the day before gets scheduled.

The hardest part was Meta's developer setup and the conversations panel. The cloud plan was expensive, so I installed it on our own server to keep costs down for my uncle. That knowledge helped me later on other projects.

Don't make anything up

I set up a panel of six simulated reviewers (three customers and three experts), and they found bugs that were losing customers: a form said “sent” without sending anything, and the recommender picked the most expensive boat under the budget. I also removed a made-up rating from the original design. Today reviews only come in through a unique link that Armando sends, and the first ones are from real customers.

We never decided to do it

The first client, from Australia, messaged us almost as soon as we posted the gig. I was slow to reply because I didn't believe anyone would write that fast, so we made the video he proposed for free. He then hired us twice.

Months without orders followed and I thought it was over. Then a client from Germany showed up. On a video call he proposed an AI video, even though I had never told him I knew how to make one. I didn't say no, and that's how our first AI video happened. Since then he has placed more than 34 orders with us.

TURMERIC, the biggest project

A playwright wanted to pitch her three-act, roughly 2.5-hour play to studios. We made her a five-minute cut of the first act for the pitch: we adapted the script, cloned the voices from a live reading, built 42 3D sets in Blender so the locations stayed consistent, and generated 21 clips. That's the method we later used for the short film.

My first website, and an ad that worked without selling

Three partners, my dad among them, started a natural products brand, but they had no online channel. I offered to work almost for free to learn. The first thing was a Shopify store: my first website and my first AI-written code, copied by hand from ChatGPT. I took the product photos with my brother, and they're still on Amazon.

I shot and edited an ad for a pancake recipe with honey, in four versions. People clicked, but nobody bought. The takeaway: the ad worked, but the store didn't convince people, because shoppers don't trust buying outside Amazon or Mercado Libre.

Not charging for what wasn't fair

On Amazon, with no sales history, things started very badly. Then I got careless while juggling high school and other projects, and advertising got more expensive. I didn't charge for the months at a loss, because they were my responsibility. I realized my pay structure was too generous for how Amazon works, and proposed switching to one based on actual profit. I also proposed not restarting sales until the inventory was in Amazon's warehouses (FBA). It worked: August 2026 was the best month in the account's history.

Using AI well

I connected the account through the API with my own scripts. When I rebuilt August, the month had 36 orders, not the 6 I had logged by hand. I also found that the discounts that seemed to be costing us were actually paid by Amazon. With the search-term data I saw where the ad money was going and reorganized the campaigns: the return ended up much better than before.

Along the way I learned to register a trademark with IMPI (Mexico's trademark office) and redesigned the labels.