Health apps track everything.
They understand almost nothing.
Most health apps are dashboards with a coat of paint. They surface numbers like resting heart rate, sleep duration and step count, and leave you to figure out what any of it means for how you actually feel. The apps that try to go further do it through generic push notifications and one-size-fits-all coaching that knows nothing specific about you.
Pulse is the other version of that. It's a fictional health companion that reads your data over time, notices patterns you'd never catch yourself, and explains them in plain language with its reasoning visible. The product is speculative. The design problem it explores is not.
I have all this data and I still don't know why I feel the way I feel. The app just shows me more graphs.
What people actually want
from health technology
I grounded the project in informal conversations with people who use health trackers daily, plus secondary research on AI adoption in consumer wellness. I wasn't asking about features. I was asking about the moments that mattered and the moments that frustrated. Four patterns shaped the design.
Who Pulse is designed for
A composite built from those conversations and secondary sources.
"I know my data is telling me something. I just can't figure out what."
- Understand why energy fluctuates
- Build habits, not streaks
- Get specific guidance
- Feel like data works for them
- Opens apps, feels overwhelmed
- AI advice feels copy-pasted
- Doesn't know which metrics matter
- Notifications feel like interruptions
- Checks data morning and night
- Has 4 health apps, uses 1
- Prefers reading to dashboards
- Skeptical of claims, receptive to data
- Thinks in systems, not scores
- Values honesty about uncertainty
- Wants to stay in control
- Would share data if trust was earned
Four principles that
shaped every decision
I committed to these before touching a screen. Every design decision in the prototype traces back to one of them.
Four screens.
One thread running through them.
Each screen answers a different design question, and the same insight runs through all four: Pulse notices that the user's 2pm energy dip is tied to their morning coffee. Following one finding from first mention to action shows how trust is built step by step. These are the live prototypes, so you can scroll inside each phone.
you slept well.
A morning briefing, not a metrics dump
How does AI earn attention without demanding it?
The home screen leads with a sentence, then one orienting score, then a single proactive insight. Everything else is there if you want it, but the story comes first.
- 1The greeting is the headline. "Good morning, you slept well" tells you how you're doing before you see a single number.
- 2The score explains itself. 76 comes with its reason: best in two weeks, because sleep and HRV both improved overnight.
- 3"Pulse noticed" card. One insight, with its reasoning right on the card: compared HRV, caffeine log and steps over 14 days. "Correlation is strong, not causal."
- 4Evidence on every card. Insights show how much data they rest on (5 nights, 9 sessions), and the log tags entries "Auto-logged by Pulse" so you can tell what you entered from what the AI did.
The transparency layer, made visual
What does it take to trust an AI's conclusion about your own body?
Tapping the caffeine insight opens the full story: the pattern, the evidence behind it, and what you could do about it.
- 1The chart makes the claim visible. The 1:45 to 3:15pm dip window is highlighted in rose so you see the pattern in a couple of seconds.
- 2"How Pulse found this." A four-step trace where every step names its source, like HealthKit HRV or the caffeine log.
- 3It tests the counter-case. Step 3 checks the days that break the pattern: on noon-coffee days the dip was 40% shallower. Looking for disconfirming evidence is what makes the conclusion believable.
- 4Offers, not actions. A 1:30pm walk reminder and a 5-day noon coffee test. Neither starts until the user picks one.
Patterns over time, already interpreted
How do you show change over a week without asking people to do analysis?
Trends is the bigger picture. Pulse writes the summary, and every chart below comes with a one-line takeaway. The dark banner separates Pulse's narrative from the raw data.
- 1The summary names the cause. Recovery is up because 6 of 7 nights fell within the same 30-minute bedtime window.
- 2One outlier, flagged everywhere. Tuesday's late night shows up in rose on the sleep chart, and the slower recovery shows up on Wednesday's HRV, with the cost spelled out: 40 fewer minutes of deep sleep.
- 3Every chart has a takeaway. Each metric card ends with a single line from Pulse, like workout days showing 18% lower stress.
- 4A month at a glance. The 28-day heatmap makes the low mid-week days stand out without reading a single number, and today's date ends the grid.
A conversation that ends in action, on the user's word
What does agentic AI look like when it acts on health data?
The user asks why they're tired at 2pm. Pulse answers with the same evidence as the insight, handles a follow-up worry with the user's own data, and only acts once they say go.
- 1"Reading:" strip. Sleep, HRV, Activity, Journal. You always know what Pulse is looking at in this conversation.
- 2Evidence inside the answer. An inline data card, plus the line "This is a pattern, not a diagnosis."
- 3Personal, not generic. Asked whether later coffee hurts sleep, Pulse checks the user's own bedtimes (11:04 vs 11:02pm) instead of quoting general advice.
- 4Acts only after "Let's do it." Then it confirms exactly what it will track, the dates, the reminders, and when results arrive on the 28th.
The reasoning trace, inside every card instead of behind a tap
Every Pulse insight shows a one-line reasoning trace inline rather than behind a "how did we get this?" button. In health, trust isn't built by the insight alone. It's built by the evidence trail.
A user who sees "Compared HRV readings, caffeine log, and step data across 14 days" can evaluate that claim. A user who sees only the conclusion has to take it on faith.
Correlation vs. causation, named explicitly in the UI
Every cross-referenced insight includes a visible qualifier: "Correlation is strong, not causal." It's a designed element, not fine print. Health AI that overstates its certainty even once can lose the user for good.
Showing the limits of what Pulse knows is a feature. It's what makes every other claim more believable.
The 5-day experiment, structured autonomy
When Pulse finds a pattern you can do something about, it proposes a structured experiment: one specific change, a defined timeframe, and a promise to track and report results. It's the most deliberate agentic choice in the project.
It gives the AI real work to do on the user's behalf while keeping the person in control of whether it happens. Pulse designs the experiment. The user decides to run it.
| Design goal | How it was addressed | Where to see it |
|---|---|---|
| Insights feel personal | Every card references the user's own time window, duration and data source | All 4 screens |
| AI reasoning is visible | A reasoning line shows inline on every Pulse surface, no tapping required | Screens 1, 2, 4 |
| User stays in control | Pulse offers; nothing starts until the user chooses it | Screens 2, 4 |
| Data anxiety reduced | Home leads with a plain-language read, not a wall of metrics | Screen 1 |
| Change over time is readable | Pulse-written summary plus a one-line takeaway on every chart | Screen 3 |
| Trust built step by step | Confidence qualifiers and "pattern, not a diagnosis" are permanent UI fixtures | Screens 1, 2, 4 |