Gonzalo Torras Serrano

Splendom Suites is a corporate-accommodation company in Madrid. My first work with them was an assistant for internal use only, for the sales team: 30+ Word documents and several spreadsheets of apartment data, turned into a single conversational interface. The first version leaned on a cloud LLM (Llama 3 70B) and invented amenities.

So I rebuilt it as an NLU/NER pipeline: RoBERTa fine-tuned to classify the intent of each query and to extract addresses, unit types and 50+ amenities (F1 0.92), a synthetic-data pipeline that grew 400 real descriptions into 10,000 training examples, deterministic filtering over a JSON inventory (no hallucinations by construction) and CPU inference on an 8 GB server so nothing leaves the company.

Then the company started growing fast, and a chatbot that only answers questions about apartments stops being enough: every new need (the mailbox, the spreadsheets, the websites) would have meant another model and another integration. It made more sense to move to Hermes, an open-source agent, again for internal use only, that I connected to the whole company: the websites, the mailbox, Drive and Sheets, the internal Kanban board and the single sign-on. It maintains the websites on its own (a new apartment goes live from a spreadsheet row), helps with marketing and new ideas, and works day to day next to the staff: it receives tasks by email or on the board, does them, assigns work and reports back with a single email.

Along the way I built the group's public websites: splendomsuites.com with 400+ apartments, splendomoutpost.com for the B2B side and elhotelinvisible.es, the group's social project.

The result: nobody at Splendom needs to be dedicated to the website any more. Hermes does that job on its own, and quite a few others.

deretiro.es lists Catholic retreats, parishes and retreat houses across Spain: more than a thousand retreats and hundreds of parishes gathered from many separate, scattered websites, where the information was never in one place, and normalised, with sister sites in German and Italian and RetStay, a booking site for 600+ retreat houses.

Rebuilding the site was the easy part. The real work is PalomIA: a process that combines AI with deterministic rules so the site updates itself. It discovers retreats in different sources, extracts and validates them under strict rules (dates, parish, title), publishes automatically several times a day and sends anything doubtful to a human review queue. It also completes what is already published and is connected to the parishes' mailbox, which it reads but never writes to.

The site grows every day without anyone editing it.

Psychology Clinic is a psychology practice in Madrid (its real name is withheld for security reasons, as agreed with them). I am building and maintaining its management platform: patient records and couple or family cases, an agenda that knows the clinic's rules, session notes, a ledger of charges and payments, an email client synced with the clinic's own mailbox, video sessions, informed consent with electronic signature, automatic reminders, roles and permissions, and an audit trail.

What used to be done by hand now runs on its own: appointment reminders, consent flows, mail sync, session bookkeeping, informative meetings for new patients.

Everything is self-hosted on servers inside the European Union. No third-party SaaS ever sees a patient's data.