Canary

Timeline

Jun 2025 - Sep 2025

Role

Product Design Intern

Scope

Consumer dashboard, six shipped views

Canary sells at-home kits that measure what a household is actually breathing. The lab was already right; nothing between the result and the family was. I designed the consumer layer that turns lab-grade metals telemetry into a decision, across one summer.

Consumer health dataDesign for comprehension

3× faster interpretation of lab results against the raw report
10+ data signals collapsed into a single consumer view
Six views shipped, each answering one question a household asks

Overview

Value proposition

Turn a lab-grade exposure report into four answers a household can reach in under a minute: am I safe, how is it changing, what is causing it, what do I do now.

Why a consumer layer

Why a consumer layer?

Every screen in this case study is a surface I created, not a restyle of one that already existed. Canary's lab pipeline was already accurate; nothing stood between that result and the family who paid for it.

The kit was sold to households and the output was written for laboratories. I designed the entire consumer layer that closed that gap.

What arrived before any of this: sixteen components, four columns of statistics, and no answer.

Pain points

Imagine

You pay for a kit because your kid keeps waking up with headaches. Two weeks later a PDF arrives: sixteen components, four columns of statistics, units you have never seen. Nothing tells you whether that is normal, bad, or urgent.

A number without a benchmark is not information. Users could read the value and still not know whether to act on it.
The report was ordered by chemistry, not by danger. Finding the worst thing meant reading all of it.
Nothing named a next step, so understanding the risk did not change anything in the house.
Ten separate signals arrived as ten separate facts rather than as one account of the home.

Users were left to decipher unfamiliar units, thresholds, and risk levels alone. Without a translation layer, environmental health data does not just go unused, it gets misread.

Features overview

Six views shipped. Each one exists because research showed a household could not answer one of the four questions from the raw report.

Exposure trends

Read against the safe limit, never alone

Plotting against the 100% limit makes escalation visible without any arithmetic.

Sensor ranking

The worst room surfaces first

Ranking by severity replaces the manual comparison users were doing by hand.

Metal thresholds

Ranked by proximity to its own limit

Every metal has a different threshold, so relative risk only reads once they are normalised.

Recommendations

Tied to the reading that triggered it

Turns an exposure insight into one specific action, today.

Weekly heatmap

By room, and by day

Surfaces the spaces that are consistently high risk rather than occasionally bad.

Risk composition

What drives it, and where it is heading

Composition beside alert volume, because overall risk needs both halves.

Process

The decisions behind the screens, and what settled each one.

My contribution

Research. Moderated interviews, think-aloud walkthroughs, task-based evaluation, and behavioral observation.

Definition. The four questions, the six design principles, and the information architecture that came out of them.

Design. Flow, all six views, the visual system, and the handoff.

Weeks 1 to 4

Research and define

Interviews and think-aloud sessions on the existing report, then the four questions.

Weeks 5 to 8

Analyze and structure

Flow mapping, the six principles, and low fidelity layouts argued against them.

Weeks 9 to 13

Design and refine

Mid fidelity, the visual system, and the six shipped views.

Research

Understand the reader, define the misreads, design the translation.

Method

I leaned on real-time cognition over recall, because people cannot reliably report how they read a number. Watching someone open the report told me more than asking them about it afterwards ever did.

Who

Who is actually opening this report
What they already believe about their home
What made them buy a kit in the first place

How

How they read it, in order, out loud
Where they stop and re-read
What they do with the number once they have it

What

What they get wrong, and how confidently
What they need before they will act
What makes them close it and do nothing

What came back

Three findings, and all three of them contradicted the instinct to show more data.

Numbers without context increase anxiety. Users felt worse after seeing results when no benchmark was provided.
Users look for what is most dangerous first, not for the full dataset. Completeness was actively unhelpful.
Actionability determines trust. Without guidance, people delayed or ignored the risk entirely.

Challenges

The report failed in three different ways at once, and each one needed a different instrument.

Numerical

Units nobody outside a lab has seen before.
Thresholds that differ for every single metal.
Ten signals arriving as ten unrelated facts.

Emotional

A health result read alone, with no one to ask.
Urgency that has to land without causing panic.
Anxiety rising with precision rather than falling.

Behavioural

Understanding a risk and still doing nothing.
No obvious next step that fits in a weekday.
No way to tell whether anything you did worked.

Ideate

Map the path, set the constraints, then draw.

Flow mapping

I mapped the whole path from the kit arriving to the household changing something, so the six views could be argued about as one journey rather than as six screens.

Result landsAm I safeHow is it changingWhat is causing itWhat do I doDid it work

The system flow, from awareness through to action.

The same path runs at the top of this page in the built product: risk first, the trend under it, and the recommendation attached to the reading that triggered it.

Design principles

Six principles, written before any interface work. Every later argument about a chart got settled against them instead of against taste.

Context before numbers

Measurements only matter against limits, comparisons, and trends.

Prioritize risk first

Surface the most dangerous thing, not the complete dataset.

Reduce cognitive load

Interpretable in seconds, not minutes.

Guide emotional response

Communicate urgency without causing panic.

Make action explicit

Every insight names the next best step.

Unify signals

Many sources should read as one story, not fragments.

Fidelity, in that order

Hierarchy got settled in grey boxes so that the visual pass was never asked to fix a structural problem.

Low fidelity, hierarchy only

Layout and reading order argued against the six principles, with nothing on screen that could win an argument by looking good.

Mid fidelity, interaction resolved

Real data volumes, real label lengths, and the interactions that the risk ranking depends on.

The product

Six views, one dashboard. The whole thing, end to end.

Every element on it exists to move someone from uncertainty to a plan, and everything above is the argument for why each one is shaped the way it is.

Reflection

Designing the first consumer surface on a working lab pipeline meant the hard problem was never the data. It was that the data was already correct and still not doing anybody any good.

What it taught me:

Context is worth as much as accuracy. Precise environmental data only becomes useful once it is grounded in benchmarks and reference points. The lab was never the bottleneck, and adding decimal places would not have moved a single household.

Action is what builds trust. Users were measurably more confident when an insight translated into what to do next. On a risk product guidance is not a nice-to-have, it is the product.

Reducing cognitive load reduces anxiety. Simplifying how information was presented improved both comprehension and follow-through. On a health product those turn out to be the same metric.

Next project

Somni

All work