My own haberdashery: five sales channels, one inventory and six applications I built to run it.
Node + TypeScript · React + Vite · Offline PWA · PostgreSQL · Docker · Mercado Libre API · Amazon SP-API · Mercado Pago · Local AI
One central inventory. Every channel reads the same stock.
Scan, count and trace every movement. Even offline.
Sales answers with local AI. Conversations stay with you.
Three years of Hilos y Encanto sales connected to see what drives the business.
Twenty-eight months brought together. Every entry keeps its origin.
Hilos y Encanto is my haberdashery. It sells over the counter in San Cristóbal de las Casas and online across Mexico: Mercado Libre, Amazon, its own web store and WhatsApp. A small team runs it, ringing up sales, packing, restocking and answering messages all day.
Before writing a single line of code, I ran it. I counted boxes, packed shipments, cancelled sales of items that were already gone and reconciled by hand what no system could. Everything I built here came from that counter, not from a requirements document.
Five places to sell. Five versions of how much was left.
Every channel kept its own count. The counter did not know what went out through Mercado Libre, the web store did not deduct stock when someone paid, and the warehouse was a count on paper. Selling an item that no longer exists is not a rare mistake: it is what happens when nobody has the right number.
And a haberdashery is the worst case: thousands of small SKUs, nearly identical colours, boxes filled with labels from different listings. No off-the-shelf software understands that operation.
Four rules I learned running the business, which the software follows without exception.
The system never deducts physical stock on its own. The team deducts the box when packing and when building a shipment; the software records, warns and compares.
An automatic deduction plus a human one is a double deduction. I would rather the system ask than believe it knows what is in a box it has never seen.
One SKU can cover several listings, each with its own physical pile. Channel and warehouse are matched by the listing code, and only by that.
I learned it by breaking production: matching by SKU switched off listings that did have stock. Since then a self-test watches for it.
«Zero in the warehouse» can mean it does not exist, it is not boxed yet or it has no label. Sellable stock is boxes plus loose pieces.
Ignoring loose pieces produced false oversells. Telling the three zeros apart avoids restocking what is already on the shelf.
The point of sale and the warehouse app check identity against the attendance app: the same person signs in everywhere with their phone or business email.
Each application still owns its permissions. A promotion in one system does not open another system's till.
The central inventory is the single source of truth. Channels read from it, the warehouse confirms it and the data tells what happened.
At the counter, on Mercado Libre, on Amazon, in the web store or over WhatsApp.
Every sale lands in a ledger that is never edited: corrections are added.
Stock is written back, and every write is read again to verify it.
The warehouse scans, packs and deducts from a phone, even with no signal.
Three years of reconstructed sales show what moves and what does not.
Six applications that talk to each other. A point of sale for the counter. A warehouse app that scans, counts by location and works offline. A central inventory connecting Mercado Libre, Amazon, the web store and the bot. The hilosyencanto.com store, which no longer keeps stock of its own. A WhatsApp agent with a CRM that answers using local AI. And attendance tracking, which is also the front door for the whole team.
On top sits the data layer: three years of sales rebuilt into an analytical model, and twenty-eight months of spreadsheets migrated into a ledger traceable to sheet and row.
Everything runs on my own server, in containers, behind a tunnel and with no ports open to the internet. Each piece has its own published case with its decisions and its numbers.
Figures are published in units and movements. The business's monetary amounts are not published, by the project's own rule.
That building for your own business is unforgiving: if the system lies, I find out at the counter that same afternoon. That pressure is what made every piece admit its limits instead of inventing a comfortable number.
And that operating experience is the advantage. The most valuable rules —a person deducts, zero has three meanings, the SKU is not enough— were in no manual. I learned them behind the counter, and they are the same ones I bring to my clients' systems.
Does your business sell in several places and nothing adds up? I already solved it in mine.
Let's talk