Local-first Notion health tracker with TRIBEv2 brain analysis, AI health insights, symptom logging, goals, medications, appointments, and a browser UI
This was supposed to be a Notion challenge submission.
I built most of it close to the deadline, got something working, and then missed the window. No big failure story. Just underestimated how long the messy parts would take.
After that, keeping it private felt pointless. So I pushed it to GitHub.
Around the same time, I came across Tribe v2. That changed how I looked at this project. Instead of treating it like a failed submission, I started treating it like something that could keep evolving in public.
That is what this is now. Not finished. Still useful.
The actual problem I was trying to solve
I sometimes already track things in Notion:
• Sleep
• Workouts
• Random notes about how I feel
The problem is not tracking. It is what happens after.
Nothing.
No aggregation. No patterns. No feedback loop. Just logs sitting there.
Every week I would think I should look at it properly. I never did.
So this project is basically me outsourcing that thinking step.
System design
The architecture is simple on paper and annoying in practice.
Pipeline
• Fetch data from Notion databases
• Normalize it into a consistent structure
• Send it to an LLM
• Write the output back into Notion
That is it. No fancy orchestration.
The difficulty is everything in between.
Notion is not a real database
At first glance, Notion feels structured. It is not.
Things that break over time:
• Property names change
• Data types shift
• Fields get added or removed
If you build with fixed schemas, your system breaks quietly.
What I did instead
I treated Notion as semi structured data:
• Map fields dynamically instead of hardcoding
• Use fallback parsing when fields do not match
• Normalize everything into an internal schema
Example internal format:
{
. "date": "2026–03–20",
. "sleep_hours": 6.5,
. "workout": "strength",
. "mood": "low"
}
No matter how messy the source is, the model only sees this cleaned version.
Data normalization is the real system
Most of the work went here.
Steps
Extract raw values from Notion API
Convert them into usable types
Handle missing or inconsistent fields
Align everything by time
Examples:
• "6 hrs" becomes 6.0
• Empty fields get dropped from inference
• Mixed labels get standardized
If this layer is weak, everything downstream gets worse.
LLM layer
The model is not used as a general assistant.
It has a narrow job:
• Summarize recent data
• Spot simple patterns
• Suggest small adjustments
Input structure
Each run includes:
• Recent data window
• Aggregated values
• Instructions that limit scope
Example:
Sleep: [6, 5.5, 7, 6]
Workout: [yes, no, yes, yes]
Mood: [low, medium, medium, high]
Task:
Identify patterns
Avoid assumptions without enough data
State uncertainty clearly
The main issue: the model guesses
Even with weak data, it tries to sound confident.
That is a problem, especially for anything health related.
What I added
• Minimum data thresholds before running inference
• Prompts that force uncertainty
• Restrictions on long term claims
• Filtering outputs that sound too certain
It still makes mistakes. It just makes fewer confident ones.
Writing results back to Notion
Outputs are stored as:
• Daily summaries
• Weekly insights
• Separate logs for traceability
Each output includes:
• Timestamp
• Data window used
• Generated insight
This makes it easier to debug and iterate.
Why I stayed inside Notion
I considered building a separate app.
That would solve a lot of problems:
• Cleaner schema
• Better validation
• Fewer edge cases
But nobody wants another health app.
Notion already has the data. So I built on top of it instead.
The tradeoff is dealing with inconsistency.
Influence from Tribe v2
This project shifted direction after I came across Tribe v2.
The main idea that stuck:
You do not wait until something feels ready.
You ship it. Then improve it in the open.
That is exactly what this repo reflects. Some parts are solid. Some are clearly not. That is fine.
What is still broken
A few things are still rough:
• Sparse data leads to weak outputs
• The model confuses correlation with causation
• Some insights sound better than they are
• No feedback loop yet to measure usefulness
The system works. It just does not always matter.
What I would change
If I rebuilt/rework this:
• Define a stricter schema earlier
• Separate ingestion and AI layers properly
• Add better logging from day one
• Focus more on actionable insights, not just observations
Where this could go
A few directions that feel real:
• Long term memory instead of short windows
• Feedback loops to track if suggestions help
• Wearable integrations
• Confidence scoring for outputs
Or it might just stay like this. A small layer that makes Notion slightly smarter.
Closing
Missing the deadline changed the trajectory of this project.
If I had submitted it, I probably would have moved on.
Instead, it is now something I can keep improving without pretending it is finished.
Right now, it is useful enough to keep using.
That is enough.
Top comments (0)