I Gave AI All My Personal Data. Here's What Happened.
Last week I shared about why I'm starting to write during the AI transition. This is the first real piece — what I actually built.
I have an AI assistant named Max. He lives in my chat. Every morning at 7:30, he sends me a voice note with my schedule, my to-dos, any urgent emails, and last night's football scores (Man City, Inter Miami, and Corinthians, in case you were wondering).
He also manages my calendar, tracks our family finances across accounts and currencies, logs my health data, plans our travel, and drafts emails for me to approve. He's not a chatbot. He's an agent — with real access to real data and real actions.
My wife Carmen has her own version. We call hers Ella. Same backbone, different identity, scoped to her data, talks to her in her chat. Both run 24/7 on a small server, gated by an approval queue whenever either of us wants them to do something that touches the outside world.
It started as a frustration project. 2k commits, 175k LoC later, it runs more of my day than I'd care to admit.
How it started
Last December I was spending hours every week on personal admin. Reconciling accounts. Checking if bills were paid. Updating spreadsheets. Replying to emails that didn't need a human touch. I work as a software engineer at Meta — I spend my days helping ship products to a lot of people. My personal life was running on my memory and my spreadsheet.
So I asked: what if I gave Claude access to my actual data?
The first version was embarrassingly simple. A Next.js app on Vercel, Supabase, a chat interface, and a couple of tools that let Claude query my calendar and todo list. I showed it to Carmen. She said: "Cool, but can it actually do anything?"
Fair point.
What it does now
Five months later, the system has a stack of modules:
- •Calendar, email, and contacts (synced from Google)
- •Chat message search and sending
- •Finances across a few accounts and currencies — bank, credit cards, investments, retirement accounts
- •Health and weight tracking
- •Travel planning with itineraries and expenses
- •Notes, todos, saved links
- •Document storage and search
- •Places, activities, blog posts (like this one)
- •A memory layer that knows what we like and what's joint vs personal
When I ask Max "how are we doing on travel spend this month?", he doesn't guess. He queries live data, pulls current rates, and gives a number. When I say "book dinner with John next Thursday at 7 at that Italian place we liked," he creates the event, pulls the address from my places list, and adds it. When Carmen tells Ella "remind me to call mum tomorrow morning," it's a todo, with context.
How it's built
The stack is straightforward:
- •Next.js 16 + Supabase (Postgres + Auth + Row Level Security)
- •Claude via the Anthropic API for the synchronous web chat
- •A CLI called `pa` that lets me — and the agents — interact with every module from the terminal
- •A chat client running on a small server (VPS)
- •Per-message ephemeral Docker containers, each with Claude Code, the conversation history, and a pre-authenticated `pa` CLI
The chat flow is the interesting part. When a message lands, the server spins up a fresh container, runs Claude inside it with full tool access, returns a reply, and kills the container. Stateless between messages except for the database.
With Opus 4.7 — released last month with hour-long autonomous runs — those containers can now handle multi-step jobs that used to need babysitting. "Check my calendar for next week and build a packing list for London" fans out to the calendar tool, the travel tool, and the todo tool. Claude figures out the order.
What broke
A lot.
The first time I gave Max access to email, he tried to reply to a spam message. I added an approval system the same day. Now any action that touches the outside world creates a pending action that one of us approves in the web app.
The financial side was the gnarliest. With accounts in different currencies, investments with live tickers, credit cards with outstanding balances — getting the totals right is harder than it looks. I once spent a Saturday chasing a significant discrepancy that turned out to be a timezone bug in how I queried exchange rates.
Context windows were a real constraint. The full financial state alone is over 100K tokens expanded. Before Anthropic's life-saver 1M context window became the standard, I had to build summarisation layers and smart querying so Claude gets just what it needs, not everything.
And voice notes. Getting Max to sound natural took iterations. Early versions opened with "Good morning, Victor. I hope this message finds you well." No. Just tell me what's on the calendar.
What surprised me
Three things.
First, how fast it became indispensable. Within two weeks I couldn't go back to checking my calendar manually. The morning brief changed my routine — I now start every day knowing exactly what's ahead without opening five different apps.
Second, the trust calibration. There's a sweet spot between "AI does everything" and "AI does nothing." The approval queue for external actions was the unlock. We let the agents read our data and suggest things. We don't let them send anything without one of us tapping yes. That single boundary made the whole thing work.
Third, how much it taught me about AI. Building a real product with tool use, long context, multi-step reasoning, and error handling taught me more in these past months than any book or class could have. The gap between "Claude can do this in a demo" and "Claude can do this reliably with my actual data" is wide, and it's where all the interesting engineering lives, at least for me.
Why I'm sharing this
It's a personal project. I'm documenting it because I think the lessons are useful, and because writing it down is the easiest way to figure out what I actually think.
If you're an engineer, this is one ground-truth view of what AI agents can do today — with real data, real failures, real constraints. If you're a builder, some of the technical details might be useful too.
And if you're just curious what it looks like when you hand your personal admin to AI — it looks like a voice note every morning from a very well-informed assistant who never sleeps, never forgets, and once tried to reply to spam.
If you're playing with the same idea, I'd love to compare notes.