The app's goal and two layer design.
As AI becomes part of everyday life, we all hope to get health information that is both scientifically grounded and personally relevant. We think about and implement this in 2 layers: Context layer and Experience layer.
- Context layer: facts, timeline, health background, wearable/lab data, etc., all in one exportable context object.
- Experience layer: research, summaries, doctor prep, feeds, and other outputs built on top of that context.
The missing layer
AI systems are getting better at text, health, biology, research, and reasoning. The bottleneck is increasingly not only model capability.
What's missing: an easy way to input, structure, visualize, edit, and carry personal health context across AI systems.
It is the small missing non-AI component around AI:
- structuring contracts,
- timeline UI,
- export formats,
- smooth UX for editing and visualization.
Existing health signals are useful but incomplete/noisy
Labs, wearables, genetics, clinical records, and trackers all matter. But they are noisy, partial, expensive, burdensome, etc..
Quantified signals alone cannot yet produce good personalization.
Natural-language facts are a good health data format
People do not remember health history as perfect structured data. That is fine.
Vague, partial, subjective-feeling descriptions can still be extremely valuable:
- “around college,”
- “worse after meals,”
- “started after moving,”
- “felt different from normal anxiety,”
- “my sleep got lighter starting around winter.”
Natural text is the native format for nuance, uncertainty, and variable granularity. LLMs make this format newly practical because they are good at processing text in high volume and connecting dots.
The core object: a good-enough personal timeline
A good-enough timeline is:
- not a perfect medical record,
- not a clinical coding system,
- but good enough to capture the most important lived context.
It should separate:
- facts from interpretations,
- timed facts from timeless background,
- user-stated observations from suspected correlations,
- personal memory from medical authority.
Facts are not interpretations
If a user says:
“I started iron in March and felt less tired later. Maybe iron helped.”
The system may extract:
- “Started iron in March.”
- “Felt less tired later.”
It should not extract:
- “Iron improved fatigue.”
Extraction AI is thin and optional
The extraction AI for adding new facts is just an input helper.
It helps turn messy text or voice into editable candidate facts. It is not the source of truth. The timeline and facts are the source of truth.
In local-only mode, users should still be able to manually create and edit the timeline without AI.
The main AI vision goes to the Experience layer
The more ambitious AI belongs above the personalization layer.
The personalization layer creates a portable context object. That object can later power:
- the app’s own platform-native AI experiences,
- personal research scientist/content feed,
- doctor prep,
- missing-info prompts,
- third-party agents,
- user-owned external AI systems.
The app should not lock the user’s health context into one AI / agent.
Portability
The Context layer should be easy to export and merge with other signals.
Natural-language facts can be combined with:
- lab data,
- wearable summaries,
- medical records,
- genetics,
- documents,
- user notes,
- AI agents.
Privacy
- Local-first by default.
- No login required for the core manual timeline.
- Export available.
- No AI features in local-only mode, because AI requires sending data out.
- Login/cloud mode is opt-in for sync and AI features.
- The platform does not actively sell data or train models on it.
- But users should assume cloud AI involves real privacy tradeoffs, including provider/API retention policies that may not be fully controlled by the app.
- Open sourcing is on the roadmap, but not the first priority unless users strongly request it.