/01Projects
Six projects. Each one explains what the problem was, what decisions I made and what changed for the client.
The theatre ticketing company that left Shopify to sell on its own system
A bilingual executive-coaching platform, with payments, administration and personal-data handling
Quote insurance from several insurers on a single screen
A coaching brand that sells, books and publishes on its own, in two languages
Point of sale and operations for a shoe-cleaning business
A data structure that serves different brands
/02Projects
Smaller or more closed projects, explained in a few lines.
Skincare online store with its own content manager (products, orders, coupons, blog and invoicing), Mercado Pago payments and 23 kinds of information.
A tool for a client to mark up changes directly on a live web page or a PDF. What sets it apart: the comments are exported in a format an AI can use to fix the code, with an MCP server so Claude Code reads them directly. Private prototype.
Transcribes, summarises and lets you chat with 1–3 hour audio: it splits the audio into 10-minute pieces processed at the same time and re-aligns the timestamps at the end.
Converts PDFs into editable text entirely in your browser, uploading nothing to any server: it recognises tables (with and without borders) and multi-column text.
/03Machine Learning
Work from Tecnológico de Monterrey, with the results that matter and the mistakes I caught along the way.
A hand-written classifier (81.5% correct against a 42.6% simple baseline), a random-forest model and a neural network that recognises traffic signs with 99.11% accuracy on 12,630 official test images. In one exercise I found that the test data looked too much like the training data, inflating the result; correcting it gave a more honest measurement.
Pair project. A model that catches 90% of positive cases, turned into a service and dashboard published in the cloud, with an explanatory model sheet and tests verifying that the service gives the same results as the analysis notebook.