AI & MLLLMData quality

Designing a Personal Data Model Around Recurring Activities

Personal data is usually organized by the application that collected it. A calendar knows about events, an activity tracker knows about movement, and another record may describe a place or a task.

For a proposed project that organizes personal activity data, I explored a different organizing idea: represent meaningful activities and recurring patterns across those traces. The retained MVP specification dates to August 2026.

From records to activities

The proposal describes normalizing personal traces into semantic activities, day archetypes, and recurring patterns.

That requires a distinction between an observation and its interpretation. A calendar event can say what was planned. A device record can show a measurement. Neither should silently become proof of what a person actually experienced.

An activity representation would need to retain the source and the uncertainty behind its label.

Why representations matter

The potential AI role is to help organize heterogeneous information into comparable concepts. Embeddings or language-model interpretation could be options, but their usefulness would depend on how well the resulting categories serve the person reviewing them.

A tidy set of day archetypes could be misleading if it hides missing records or forces unusual days into an unsuitable category. The interface would need a way to inspect and correct the interpretation.

A specification before an application

This is a proposed product model, not a completed MVP or an implemented embedding pipeline.

The next useful prototype would use synthetic traces and a small, explicit set of activity definitions. I would first test whether the representation is understandable and correctable before adding more automated inference.

The broader connection to my AI work is the importance of meaningful intermediate data. A model can help produce a label, but a useful personal-data tool needs to preserve what that label came from and let the user disagree with it.