You can try the main ones in the lab, with made-up data. The rest are below, in short.
Main projects
Own product · running · 2026
Ada: marketing for my channels with n8n and AI
Prepares, reviews and publishes the videos and posts of five channels. A weekly cycle, a nightly check and a queue where nothing goes out without review.
In the lab: Be Ada for a week: stories, a used-up quota, the 8 pm review and publishing.
Click any of them to see the problem it solved and how I did it
No projects of this type yet. If you have one in mind, tell me about it. Ask for a quote →
Own product · running · 2026
Ada: marketing for my channels with n8n and AI
Problem
Running YouTube, Instagram, TikTok and LinkedIn with videos, covers and copy every day eats up the evenings, and the day something breaks, nothing gets published.
What I did
Ada, a system in n8n that works in two cycles. Every week it asks ChatGPT for the stories and prepares episodes, reels, posts and campaigns; every night it looks back at the day and, if it is worth it, prepares a news video. ChatGPT writes the stories and makes the images, Google AI Studio does the voices, and editing, clips and covers come from my own video engine. Everything goes into a queue: the evening before, at 8 pm, it checks cover, copy, time, a maximum of three posts a day per network and no repeated topic within 14 days, and sends an email with a “Stop” button for each post.
Result
Approved posts go out through Buffer and the YouTube API, and the video is archived on its own. An error and alert log: if the image or voice quota runs out, it records the time and picks the work up again when there is quota; if a network fails, it asks for a manual upload. Review and alerts are live; automatic publishing is in testing.
n8n · OpenAI API · Google AI Studio · Buffer · YouTube API · rclone
Club Byte: technology explained to children aged 8 to 12
Problem
There is hardly any content in Spanish that explains to children how computing works through stories and characters, without selling a course behind it.
What I did
A YouTube channel and a blog with two main characters, Lía and Teo, and Bit, a robot made in Blender. Each episode explains one idea (sensors, conditionals, binary, algorithms) and ends with an unplugged challenge for home or class.
Result
A 3D animated intro, scripts, voices and the first episode, plus a blog for families and teachers with activity sheets. No cookies or tracking, as a children’s site should be.
Blender · Python · Educational scripts · SEO for families and teachers
Every application was deployed by hand on the server, each with its domain, login and backups done a different way.
What I did
A system that gives each app a test environment and a production one on the same server. An app plugs in with a manifest file and the system creates its route, its Keycloak user group, its domain with a certificate and its deployment from the GitHub image.
Result
About ten applications living on one server, “Deploy”, “Promote to PRO” and “Roll back” buttons, automatic rollback for the apps the server deploys, and an agent that updates itself when the tests are green.
In a small factory, orders, stock and invoices usually live in different programs or in Excel, and the shop floor stays outside the ERP.
What I did
An ERP with customers, quotes, orders, routings, delivery notes, invoices ready for Spain’s VERI*FACTU and a warehouse with batches. It has a desktop version, connects to the shop floor through Efesto and to sales through Brokkr, receives support hours from Odin and builds its PDFs with Report Server.
Result
Running (version 0.2) and open source under the AGPL-3.0 licence. Invoices are immutable and ready for VERI*FACTU; it does not issue real invoices to third parties yet.
PostgreSQL (PGlite) · Electron · OPC UA · REST API · Docker
Machine status, good and bad parts and alarms stay inside each PLC or SCADA, far from the ERP, and connecting them usually means opening ports into the factory.
What I did
A read-only shop-floor agent with 16 connectors: OPC UA, Modbus, Siemens S7, Rockwell, Beckhoff, Omron, MQTT, MTConnect, BACnet, historian SQL, files and scales, among others. The connection goes out from the factory, stores and resends if the network drops, and no plant addresses or passwords ever leave it.
Result
Tested with 22 simulated machines: 106 out of 106 data points reach Sindri. If one machine goes down, the rest carry on. Not yet connected to a real factory.
Brokkr: the CRM for a small manufacturer’s sales rep
The in-house CRM, in progress. It is a Sindri module sharing the same customers, quotes, orders and invoices, so nothing is duplicated. The goal: log a visit in under 30 seconds from a phone or handheld, even without coverage; opportunities that win themselves when the order is confirmed; and one timeline per customer with calls, quotes, invoices, payments and Odin tickets. The pipeline, activities and quotes already work inside Sindri; before going further I analysed 15 open-source CRMs.
28 example websites for made-up businesses: hospitality, beauty, sport, health, pets, retail and services. Each one has its main action (book a table, make an appointment, place an order or ask for a quote) and you can try it end to end without sending anything. Mobile first, with self-hosted fonts instead of Google Fonts and no inline styles or scripts, so they pass a strict security policy. They include a legal notice and say clearly that the business is made up. The catalogue filters by sector and comes from the same generator I use for private client demos.
HTML and CSS · Dependency-free JavaScript · Strict CSP · Own generator
An assistant that looks after a business’s customers on WhatsApp: it answers the usual questions, gives opening hours and services, helps book an appointment or place an order and hands the chat to a person when needed. Official WhatsApp Business API only. In development.
Tickets were noted down by hand, hours lived in a spreadsheet and invoices somewhere else. Nothing was connected.
What I did
My own web application where each ticket carries its follow-up, its documentation and its timed work sessions. It started on top of MantisBT and now has its own tables, with the history migrated.
Result
Single sign-on with Keycloak, every open ticket creates its task in the task manager, and pending hours flow to Sindri on their own to be invoiced. Branded email through the Mail service.
Mail and Report Server: email and PDF for every app
Two services shared by all my applications. Mail sends Arce Data Lab branded email through Resend, from soporte@arcedatalab.com, with blocks (title, button, table, image), attachments, per-app keys and a sent log. Report Server designs PDF templates in the browser with pdfme, connects to MariaDB in read-only mode and generates invoices, orders and delivery notes through an API.
The warehouse needed to see live stock from a website, but the databases sit on the internal network and no inbound ports could be opened.
What I did
A gateway in the style of the Power BI one: an agent inside the network opens a WebSocket tunnel to the website, queries MariaDB and SQL Server on demand and returns the data through that same tunnel.
Result
Up-to-the-minute stock on touch screens on the shop floor, permissions by department, an interface in Spanish, Chinese and English, and export to Excel.
A business that manages patient appointments in hospitals was working with a spreadsheet-based application that had fallen short, and hospitals do not always have coverage.
What I did
I rewrote it as a web application with its own database and a tablet app that works offline: it saves locally, queues the changes and syncs them as versioned events when the network comes back.
Result
Schedule, patients, services, payments, documents and alerts in one place, with a conflict screen to decide when two devices change the same thing. Deployed on my own server.
Business questions about the ERP (“how much have we sold to this customer this year?”) ended up in a report or an email to whoever knew how to pull the figure.
What I did
An assistant that answers by querying the ERP in read-only mode. Each type of query is a closed function: the model chooses which one to use and with which parameters, so it does not make data up. It relies on editable business notes that explain how each figure is calculated.
Result
In use at Grupo SPSIL since 2026. It answers in Spanish, English and Chinese, as text, a table or Excel, and cannot change anything in the ERP.
A network change cut off the data sources of the Power BI reports for management, sales and purchasing.
What I did
I migrated the processes to Dataflows and a Lakehouse in Microsoft Fabric, on top of A3 ERP. Every change to a view ships with its backup, its rollback script and a before-and-after check.
Result
The sales, collections and stock reports are still in use on Fabric (2025–2026), with the figures checked at every change. They are the basis of the analysis packs.
A personal application for losing weight sensibly: it generates the weekly plan from a catalogue of more than 500 foods adjusted to the calorie target, builds the shopping list from what was chosen and learns from what you mark as “liked it”. Every design decision is kept in a changelog.
A desktop application that generates a report of quotes by customer and product family straight from SQL Server or the Lakehouse. The calculation runs in the database so the computer is not loaded, with filters, export to Excel and automated tests.
A shop for a screen-printing business: catalogue, cart and checkout, with a custom admin panel to manage products and orders. Everything starts with a single command thanks to Docker.
Internal tools for two production plants: an in-house helpdesk system for support, time tracking, IT cost monitoring and integration of invoicing between the group’s plants.
A web app to organise a team’s work: Kanban and Scrum boards, calendar and agenda, wiki, notes, whiteboard, SLA, person-to-person notes, activity walls and an API designed for AI agents. It is a base version with no company data: branding, login (local or Keycloak) and optional modules are set in one file, and it installs with Docker and demo data.
A theme generator for Power BI in the browser: you pick your brand colour, see the palette on a live dashboard, check the contrast and export the theme.json.
An interactive page for understanding the most powerful and most confusing function in DAX: you change the filter context and see, row by row, what the measure counts and what CALCULATE modifies.
Four steps and four prompts for an AI to read your model and draw its dependency graph: which measures hold it up, which ones nobody uses and which connections cannot be seen at a glance.
A template for a team to define how AI is used when reviewing or building projects: a set of rules nobody can skip and two improvement loops that learn from each real case.
A way of working with AI as a collaborator rather than a search engine: framing, iteration and wrap-up, with templates designed for DAX, Power Query and documentation.
A Power BI report with sample data to decide which customers and areas the sales team should spend its time on, with a Python script that regenerates the data.
In many plants OEE is worked out by hand in Excel once a month: by the time the number looks bad nobody remembers what happened, and each supervisor logs stoppages their own way.
What I did
A complete pipeline with made-up data: three simulated machines publish over MQTT whether they are running or stopped, the reason and the good and bad parts. Apache NiFi 2 reads it, validates the JSON against a schema, applies the business rules, adds the shift and the machine name from the PostgreSQL master table and stores it. Bad messages go to an error queue, each with its reason.
Result
OEE (availability × performance × quality) per machine and shift, downtime by reason and scrap come out of SQL views ready for Power BI. Everything starts with a single docker compose up in about two minutes and, in testing, none of the 12,960 history messages was lost. It includes the NiFi flow template in JSON and a guide to scaling NiFi: back pressure, queue prioritization and when to move to a cluster.
Apache NiFi 2 · MQTT · PostgreSQL · Power BI · Python · Docker
A step-by-step manual, in Spanish, for installing NiFi on Windows and building a first pipeline in twenty minutes, aimed at analysts coming from Power BI and Power Query.
