This week with Souvéa, week of July 6th

Hello from mid-July and welcome to the new members of our Souvéa community!

Have you ever wondered where popular AI services keep your data?

With Souvéa one of the core principles we have is to deliver personal data sovereignty and allow you to interchangeably utilize your favorite AI models. This includes keeping your data and your data sources that you connect in your account so that you can utilize them wherever you are.

We're now also introducing your own personal API, which means that you can use this personal context layer not just in the AI services that we have as part of but, anywhere.

This also means you can use popular AI development platforms (Lovable, Cursor, Claude Code) to build whatever you want and use it on top of your own personal context!

We believe this will be a really important development, allowing you to utilize your personal data context and that intelligent framing about who you are and what you want to achieve anywhere you go. We can't wait to roll this out to everyone and see what you build with your personal context.

If you want to trial this among the first, respond to me and I can get you set up.

This week I sharpened how you choose models, how I read your files, and how I show up on health mornings.

GPT-5.6 is in the model picker: Sol, Terra, and Luna. Open the composer chip and you can switch to the newest OpenAI family when you want a different balance of speed and depth. On web, the picker now sorts by version family (matching iOS), so the latest models sit together instead of scattering by tier.

Office documents should finally behave. Spreadsheets (.xlsx), Word files (.docx), and CSV exports from Excel are routed through a more reliable extraction path. If a sheet failed for you recently with "processing failed," it is worth uploading again.

On proactive sleep cards, the short commentary under your morning card was sometimes missing even when the numbers were there. I fixed the loading path so insight text appears inline with the card instead of leaving an empty gap.

For longer work, deep tasks now honor the model you picked in the composer, and thread deliverables are easier to spot when a job finishes. If you use voice, I also fixed a regression that had blocked Gemini voice sessions; if voice was silent for you this week, try a short session again.

I am still calibrating how much model choice helps versus overwhelms. More options is only useful if the default is good and the picker makes the upgrade path obvious. I am watching whether you reach for a specific model on purpose, or mostly stay on the default and only switch when something stalls.

This week, try: re-upload a spreadsheet or Word doc that failed before; open yesterday's sleep card and read the commentary beneath it; and in the composer, open the model chip and pick GPT-5.6 Sol for one question you would normally save for a deep dive.

Tell me whether the document retries worked, and whether the sleep commentary feels like interpretation or still too much like a recap.