Inventory and sales systems, WhatsApp agents, delivery routing and robots that test your store. I run everything in my own business first, and every case study tells you the truth: demo, pilot or production.
David Bajonero· I reply personally, on LinkedIn $ npm run qa:checkout▸ starting test copy of the store… ok✓ checkout desktop · 10/10✓ checkout mobile · 10/10⚠ injected defect in "total" → caught20/20 green · 65 s · video + trace + summary✓ EE-02 retail: 1 to 10 per SKU, duplicates merged✓ EE-04 shipping: Bogotá, national, free threshold✓ EE-05 contact and consent validated on the server▸ video: runs/checkout.webm▸ trace: trace.zip▸ summary: resumen-es.mdexit 0synthetic fixture · no real payments
Illustrative example with fictional data. The six steps are the real ones.
An owner writes what hurts, in their own words. That is the whole brief at this point.
A 30 to 45 minute conversation about how the business actually runs: purchases, counts, who closes the register.
Lost Saturday sales plus emergency purchases. If there is no number, there is no project.
Built on the shape of their business, with fictional data. The owner sees it work before paying anything.
The system has to pass its own checks: every sale deducts, every alert fires, nothing is billed on a promise.
A sale comes in, the flour goes down, the alert reaches WhatsApp before the lunch rush. Measured in a pilot, one location first.
illustrative · fictional data
Animated illustration of the Restocore point of sale: three products go into the cart, the payment is Nequi, the sale is registered, a receipt slides out, the ingredients of the recipe drop in the location stock and the dashboard shows sales per location
demo · fictional dataA system that knows what you have at every location, deducts ingredients per recipe and hands your accountant a clean file.
Around 200 deliveries a day. A custom engine on Bogotá's real street map decides which vehicle goes where, and in what order.
The agent runs the first conversation, flags what's urgent and hands the lawyer a ready-to-use brief.
Everything I offer, I test first in a business that actually sells.
Three interfaces recreated in code so you can see how they feel. Experiments and prototypes, labeled as such.
Sofía assigns the task to Tomás with Sonnet at medium effort, Tomás delivers a file, Camila reviews it and the publish step waits in David's inbox.
A pixel-art office where Sofía, the manager agent, reads the task queue, assigns each task to a real agent and explains why she picked that model. Nothing publishes, pays or deletes without the owner's yes.
See it in the lab→A user runs /arranque, Atajo asks one question at a time, saves the three priorities, parks an idea that serves none of them, delegates by difficulty and stops before publishing.
Your personal AI operating system in Spanish for Claude Code. Free, MIT. It remembers where you were, keeps you on your three priorities, delegates by difficulty and never publishes without your yes.
See it in the lab→The camera reads the hand as 21 points in the browser, the points form the letter L and the game confirms the match without sending any video.
Learn Colombian Sign Language with your camera. The hand becomes 21 points inside the browser, so video never leaves the computer. At the Claude Code Build Day I added the flow where a signer records a word and anyone can practice it.
See it in the lab→Five kinds of systems. They all start the same way: one concrete problem, one number we can measure, and a human who signs off on what matters.
You don't know what's on the shelf at each location. Closing the register is manual, and your accountant gets the numbers late and messy.
Timeline: A demo on your case in about a week. A 30-day pilot at one location before rolling out. See the case: Restocore→The same questions arrive at all hours. Someone answers them one by one, and the urgent cases get buried.
Timeline: A demo with fictional cases in one or two weeks. A measured pilot before opening it to all your customers. See the case: Customer-facing agents→Hours of manual work that repeat every day: planning routes, keying in invoices, reviewing documents, building reports.
Timeline: Diagnosis and demo in one or two weeks. Savings measured over 2 to 4 weeks with your data. See the case: Logistics and automation→A small change breaks checkout or a form, and nobody notices until a customer complains. Or quietly stops buying.
Timeline: First flow running in one or two weeks, depending on the flow. See the case: Quality and testing→Most owners who want AI picture a chatbot. Start with the number that hurts and your data instead. A practical playbook from selling to, and building for, restaurants and shops in Latin America.
Read→In September 2026 I moved from no-code tools to shipping real software with Claude Code. Within weeks I had a 396-test inventory system, a routing engine on Bogotá's real map and a QA robot. Here's my operating model, and the mistakes that cost me.
Read→A harness is the frame around AI work: rules, limits, tests and evidence. Here's a plain-English definition, how I built one to test my store's checkout, and why it won.
Read→It depends on the problem and the value it recovers, so I don't publish prices. First a diagnosis, then a demo, then a proposal with three options, each with a setup fee and a monthly fee. You know the monthly running cost from day one.
No. The system works alongside what you already use. Your certified provider keeps handling e-invoicing. I connect, import and organize the data.
One installation handles all the locations of one business. Each location keeps its own stock, daily close and ledger, and each user only sees the locations assigned to them: the cashier at one location cannot see or change another. Catalog, recipes and suppliers are shared. Separate companies get separate installations.
Because that's the truth. A demo shows what can be built and how fast. A pilot measures results with real data. Production is what's already in use. I'd rather say it plainly than oversell.
Tell me in one message. I'll reply with a concrete idea of what can be automated, or an honest no.
I reply personally, usually within one or two business days. · Bogotá, Colombia (UTC-5)