A continuously updated index of Portugal's property and used-car markets, around 273,000 live listings (property and cars) refreshed by a self-healing scraping fleet.

On top of the data sits statistical deal-finding that scores every listing against its true comparables, valuation models, rental-yield maths, and a demand-velocity engine that answers what is actually selling, and how fast.

It even spots cross-border import arbitrage: cars that are meaningfully cheaper to buy abroad and bring in.

Buying a home or a car here is guesswork dressed up as a marketplace. I wanted to hand a regular buyer the edge that funds and dealers keep for themselves.

The live dashboards

A mesh of scraper nodes across several countries, wired together over Tailscale, pulling work from a central claim-queue so no two nodes ever step on each other.

Each source gets its own anti-bot strategy (pacing, fingerprints, rotation), and when a node trips a block or falls over, the fleet notices, re-queues the work, and recovers without me touching it.

Scraping at this scale breaks constantly. The only way to keep 227k listings fresh is a system that expects failure and heals around it.

Architecture and fleet view

n8n orchestrates the whole pipeline: it sources topics, writes scripts, renders fully designed videos in Remotion, voices them with GPU-accelerated TTS, captions them, and publishes daily across 29 channels.

No timeline, no editor, no render farm rented by the hour. Just a graph that runs itself and a GPU that earns its keep overnight.

I wanted to prove that a serious, multi-channel content operation could run with a team of zero, and learn where automation's ceiling actually is.

Pipeline and sample output

Ralph lives in Telegram. Behind the chat is an MCP tool layer that lets it actually do things: reach into a 30-service homelab, run jobs, fetch data, change state.

It routes intelligently across Claude, Gemini and a local Ollama model depending on the task, the cost and the privacy, so the right brain answers the right question.

I did not want a dashboard for every service I run. I wanted one interface I could text from anywhere, and have it just handle things.

Ralph in conversation

Drop in a Google-Maps URL for a business. Claude analyses who they are, what they do and how they should look, then the pipeline generates a complete website, deploys it, and configures DNS and email.

End to end, in under three minutes, with no human in the loop.

Small local businesses deserve a real web presence without the agency runaround or the price tag. This collapses weeks of work into minutes.

A generated demo site

A private operator dashboard that turns the raw market into decisions: which models are in demand, how fast they sell, and where a car lands cheaper imported than bought locally.

It is the signal a dealer never had before. Built hand in hand with someone who buys and sells cars for a living, so it answers the questions he actually asks.

A dealer I spoke to bought purely on price and brand, with no read on demand. The gap was obvious: give him the same market edge the platform gives buyers.

The operator terminal

A complete promotional-products catalogue, migrated off WordPress and WooCommerce and rebuilt as a fast static Astro storefront backed by a FastAPI and SQLite service.

It imports the supplier catalogue natively (Makito), runs a quote-request flow and an admin behind Cloudflare Access, and ships a Discord bot, all security-hardened (CSP, XSS fixes, a Turnstile-gated form).

A real business was stuck on a slow, fragile WooCommerce stack. I moved it to something fast, cheap to run, and fully under control.

The storefront

A Discord bot that builds its own community: one setup command provisions the whole server, a honeypot, and per-bet-type premium channels.

It is fed by my football model, posts picks on a schedule, and sells premium access through Stripe, granting the role automatically on payment.

I wanted to watch a content product run end to end with zero manual ops: provisioning, posting, paywall and billing, all automated.

The bot in the server

A plate lookup for a car dealership: enter a Portuguese registration and it resolves the make, the VIN where available, and what it can decode about the vehicle.

Built honestly around the limited public sources: it surfaces what exists and is clear about what it cannot get without paid VIN data.

A dealer kept looking plates up by hand. This turns a repetitive chore into one box and a button.

The lookup

A clean player over the Radio Garden catalogue: browse by place, hit play, and see now-playing plus auto-fetched lyrics.

It even hooks into Lidarr, so a track you like can be queued straight to your own music library.

A small, joyful build. Sometimes you make a thing just because it would be nice to have.

The player

A 20-page site for Grupo Partilha de Vivencias, a Portuguese community association, moved off a slow WordPress install to a static Astro site on Cloudflare Pages.

Donation tiers, projects, events and contact, with bot-protected forms, rebuilt to load instantly and cost almost nothing to run.

My mother is a social worker who founded GPVPT, and I have helped build and run it alongside her for years. This is part of that.

The site

A Proxmox host running around 30 Docker services, exposed to the world through Cloudflare Tunnels across three zones, with not a single inbound port open. Tailscale handles the private side.

It is also where my custom Claude MCP servers live, giving every other project a secure, tunnelled home.

I wanted production-grade infrastructure on my own terms, private, secure, and mine, for roughly the cost of electricity.

Topology and service map