Live projects and open case studies. You can open and check everything yourself.
AI agents & automation
Our own system and small demos you can check. The demos use made-up data; the code and the tests are real.
The system our studio runs on. Claude Code is the main agent. Hooks check every action, skills handle repeat tasks, a shared memory carries decisions from one session to the next, MCP tools connect it to other models and services, and a budget guard blocks paid calls above a set limit. It runs on our own machine in Nice.
About CastelOS →
One AI model checking another misses the same things. This open-source tool sends each answer to several judge models from different labs, takes a vote and sends split cases to an arbiter. In a real run on 16 code answers, 6 of them with planted bugs, Mistral, Gemini, Llama and GLM caught all 6 and passed all 10 good ones. The test set was easy: on real data the judges disagree more often.
Code on GitHub →
Six invoices from three suppliers, each with its own layout (English, French, German labels), become one table. The totals are checked: net plus VAT must equal the total. Nothing is guessed: in the demo, 4 rows come out clean and 2 are flagged for a person to review.
A daily job takes exchange rates from a public API, adds them to a table (CSV or Google Sheets) and sends a Telegram message only when a rate moves more than your threshold. Running it twice adds no duplicate rows; if the API is down, it tries 4 times and then stops with a clear message. 8 tests, no network needed. The Google Sheets path is tested on a stand-in sheet.
A messy customer export becomes a clean file. Every change is logged: what it was, what it became and which rule did it. In the demo, 39 rows go in and 36 come out, with 171 logged edits and 6 rows flagged for your decision.
PORTFOLIO
The site shows what the studio can actually build. It maps public work and keeps private work private.
Pilot projects
CastelOS runs AI on your own computer. It handles notes, automation, media, reports and workflows you control.
Evidence: You can see design notes, system status and a live demo if you ask.
Evidence ->Public channel, running
A personal news radar. It reads hundreds of sources in many languages. It keeps only what fits a narrow set of interests. It publishes the analysis on a schedule.
Evidence: Open the channel and read the output. That is the proof.
Telegram ->Public channel, running
The system runs a regional news channel. It scans sources, drafts items and publishes them on a schedule. It works in French and English.
Evidence: Open the channel and read the output. That is the proof.
Telegram ->Private product line
A voice companion. It handles calls and remembers details. It is designed for care.
Evidence: Real work in progress. Details stay private.
Evidence ->Product prototype
An AI tutor. It covers lessons, voice practice and pronunciation feedback.
Evidence: A working prototype — demo on request.
Evidence ->Open-source tool
A small tool that checks graphics card memory before an AI model loads. It helps stop your computer from freezing.
Evidence: Public GitHub repository.
GitHub ->Public map
A curated list of tools for running AI on your own computer. It helps you choose tools that keep data on your machine.
Evidence: Public GitHub repository.
GitHub ->TLX
An AI media channel that runs by itself. It publishes around the clock and fixes its own errors.
Anna
A voice companion that calls, listens and remembers. It is built for elder care.
French Tutor
A working prototype of an AI tutor with lessons, voice practice and pronunciation feedback.
Self-heal
A system that finds and fixes its own failures. It runs live in real use.
No stock logos. No invented testimonials. These systems are real and running.
gpu-memory-guard ↗
A small open-source tool. It checks graphics card memory before an AI model loads. It helps stop your computer from freezing. It is MIT-licensed and used on our own computer.
awesome-local-ai ↗
A curated list of tools for running AI on your own computer. It is the map we use ourselves. We maintain it in the open.
castelos-public ↗
Public build notes for CastelOS. You can read the principles, design notes and updates as the system grows.
We set the task, documents and limits before work begins.
An important step stays visible and a person approves it.
You get a file, note or working result you can keep.
This is how pilot projects work. You give clear input, stay in control, and get an independent result.
One person reads this inbox — the same person who builds what you order.
Tell us what you have and what result you need — even a single small task. No brief required: within 48 hours you get a plan and a price in writing, a question, or an honest “not a fit”.
Prefer a call or a messenger? Write or call — the studio's founder answers.
Phone · WhatsApp · Telegram
A one-person AI studio in Nice. Whoever answers your message is the one who does the work — no account managers in between. Files are processed on our own machine, not in someone else's cloud.
Castel Studio assistant
virtual assistant · replies instantly
Hello! I’m the virtual assistant of Castel Studio. Ask about services, prices or how we work — I answer from the site. A human, the studio’s founder, replies through the form.
The conversation is kept for 30 days and processed by a local model on our computer in the EU, with no third-party AI service. Details