Case study · June 2026 to now
Vincenzo
A personal AI command centre, built into my operating system.
Buonasera, Raoul.
Everything is under control. Hold Page Down to talk.
What he is
Vincenzo is a local AI assistant that lives inside my desktop. He talks and listens, remembers what matters, plans my day through Taccuino, runs deep research and deploys specialist agents in sandboxes. Nothing privileged happens without my approval.
He has a character too, a retired Italian cellar master and consigliere. Warm, precise, discreet, with a dry wit. When nothing needs me, he says so. "Nothing needs you. Bene."
I designed and directed Vincenzo through AI-assisted development, with ChatGPT as planner and reviewer, Codex for scoped jobs and Claude Code for the large builds. The architecture, design, specifications, testing and diagnosis are mine.
Start to present
Five months, one bottle at a time.
June 2026
The first sketch
The information I worked with every day lived in spreadsheets, notes, tabs and chat windows that never talked to each other. VinDeck was my first answer, a single command deck where an assistant could see the whole picture.
September 2026
A face on the desktop
I rebuilt the idea natively on Omarchy, my Arch Linux desktop. First a tracker, then the foundation, then a bar panel with a voice orb. By 2 September he could hold a typed conversation, off by default and capped on cost.
October 2026
Agents, carefully
The first agent could only read a frozen snapshot and propose. Once that was proven, one overnight build added an audit ledger, work orders, a Second Brain, local voice and eleven specialist roles.
Late October 2026
The deck
On 31 October I approved the full-screen deck, a native Qt app in the Osaka Jade theme of my desktop. Claude became his brain, running without tools, so the model suggests and the system decides. On 2 November I sealed version 0.4.0.
November 2026
Research and a manager
A Research tab arrived, where I can ask him to look into anything and talk it through when the packet is in. Aldo, a manager agent, now runs longer campaigns and asks one plain yes or no at a time.
How it's built
The bottle method
I build Vincenzo like a winemaker, in bottles. Each one is a small, named piece of the system, sealed once it is proven.
- 01 Inspect Read what is really on the machine.
- 02 Plan Write down exactly what changes, and what will not.
- 03 Approve I say yes to that plan, explicitly.
- 04 Implement Build it in a candidate copy, never on top of a sealed one.
- 05 Test Automated suites run. The 0.4.0 seal ran over 450 tests.
- 06 Retest I use it myself, live. Only my word counts as acceptance.
- 07 Checkpoint A frozen copy with checksums, results and recovery steps.
- 08 Seal Archived and never edited again. It doubles as the rollback.
Sealed so far Tracker · Bottles 1A, 1B, 2, 3A · Vintage 0.4.0 "Voice and Persona"
Inside the deck
What each tab does
Real renders of the deck, with sample data for a creative project.
01 · Ctrl+1
Home
The orb, his state in words and the conversation. Replies that need permission arrive with a work-order card to review or approve.
Hold Page Down and talk.
Hold one key anywhere on the desktop, speak, let go, and he answers aloud. Hands-free mode listens again after he replies, and only his name interrupts him.
Speech-to-text runs locally with Whisper, and he speaks through Kokoro as Nicola, an Italian voice speaking English, streamed sentence by sentence. The orb is rebuilt here from the deck. Try its states.
02 · Ctrl+2
Today
My day, read from Taccuino. Vincenzo's read on the day, the schedule, tasks, habits and a focus timer.
03 · Ctrl+3
Agents
Who's working, step by step. Each job shows its chain of roles, timeline, grants, sandbox and budget, with a cancel button.
04 · Ctrl+4
Work orders
An approvals inbox. Each envelope reads like a document, objective, what it may do, the plan, limits and rollback. Approving binds that exact plan.
05 · Ctrl+5
Memory
Scoped search of the Second Brain. Agents never get the whole vault, only small recorded slices of context.
06 · Ctrl+6
Activity
Everything he did, from a ledger that can verify its own chain. Every finished job leaves a receipt.
Agents at work
One approval, a whole team.
A sample work order for a creative project. Vincenzo drafts the envelope, I approve it once, and specialists take it from there inside a sandbox.
- Objective
- Prototype a 3-surface projection map in a sandbox and record a test render.
- What it may do
- file_readlocal_compute
- Produces
- prototype.diff · test-render.mp4 · run receipt
- Rollback
- Nothing is applied until I approve the diff. The sandbox is deleted afterwards.
Approved · bound to digest 2f70c9b1…
Aldo · manager agent
Moodboard research is ready. 24 references from 3 lanes, 3 dropped for weak sources. Draft the style guide?
A bare yes or no answers him directly, with no model call.
A small segue
Scheduling, through Taccuino.
Taccuino is my own notebook app, with pages, boards, a calendar and habits. It stands on its own, but it is woven into Vincenzo.
Small changes happen at once and Taccuino lets me know. Anything bigger becomes a plan that waits for my approval, and a backup is taken first. Try it.
Vincenzo proposes
Lay out the gallery install week- Studio stand-up
- Projection test at the gallery
- Edit reel v2
- Reply to curator
- Run · streak 4
Vincenzo updated your notebook
Trust and safety
The brain talks. The system acts.
The model can only name an action from a fixed list. Code carries it out through the same gated paths, and nothing is ever approved by the model.
Local compute and the job's own workspace. Starts on its own.
Files, the vault, network, providers, the code graph. Waits for my approval.
Microphone, provider calls, headless browser. Needs a switch in Settings as well.
Payments, top-ups, subscriptions, purchases. Never automated.
- SandboxesAgent jobs run isolated, with no network unless granted.
- AuditEvery decision goes into a hash-chained ledger, secrets redacted, audio never stored.
- Scoped memoryAgents get small recorded slices of context, never the whole vault.
- Web pages are dataAfter reading the web, only more reading may follow in that turn, never a change.
- Nothing sent for meNothing is sent, submitted or bought without my explicit word.
Where it stands
The core is sealed. The newest agents are in testing.
Tracker, Bottles 1A to 3A and Vintage 0.4.0, including the native deck, voice, persona, Taccuino control and memory.
Research, Aldo, agent campaigns and model routing that starts with Sonnet and escalates to Opus.
Accept and seal 0.5.0, then a weekly review view. Write-capable coding agents stay unauthorised until they earn it.
