Case study
My tools were organized. My life wasn't.
Why I started building LLM4LIFE, and what it taught me about connecting AI to real systems without losing control of the data.
I've tried a lot of productivity tools.
The frustrating part wasn't that any one of them was terrible. It was that each one knew only a small piece of what I was doing.
A task in one place. A calendar event somewhere else. Notes in another app. Project work in GitHub. And then an AI assistant that could discuss all of it but didn't necessarily know which system should be updated.
I wanted to fix that.
That's how LLM4LIFE started: not as another chatbot, but as a way to connect the tools I already use.
The first mistake would have been putting everything in one place
My original temptation was to build one large system that owned everything.
But that creates a different problem. Calendars already know how to manage time. GitHub knows code. Notes are useful precisely because they aren't always rigid records.
I didn't need to replace every tool. I needed to decide which tool should own which information.
For example, the current architecture treats PostgreSQL as the durable home for machine-readable actions and state. Google Tasks is a human-facing view of actions. Google Calendar handles scheduled commitments. Notion can serve as a view or migration fallback without quietly becoming a competing source of truth.
It sounds like a small distinction. In practice, it changes how every integration behaves.
AI should coordinate the work, not invent the state
The important question isn't “can a model create a task?”
It's “what happens if the model creates the task twice, writes to the wrong app, or loses its connection halfway through?”
I started putting more effort into explicit ownership rules, stable identities, authenticated integrations, reconciliation and rollback paths.
The repo's runtime-status notes document live migrations and checks, including moving action state into Neon and preserving contact identities through a provider migration. Those notes also identify unfinished verification work. I think that honesty is important: a migration isn't complete because a script ran once.
Why this matters to me
I like AI tools. I use them a lot. But I don't want a system that feels clever in a demo and confusing the moment something fails.
I'd rather have a smaller automation I can inspect, reverse and trust.
LLM4LIFE is still evolving, and I wouldn't pretend every part of its roadmap is finished. What it has given me is a clearer way to think about AI systems: models can help make decisions, but the underlying data needs rules that don't depend on the model getting everything right.