We actually talked to people. (Wild concept.)
The first time I tackled Canvas was in my HCI Methods course. Our team , Laney, Ian, Lilly, and myself , picked UCSC undergrads as our community. Not exactly a hard sell. We all knew Canvas was a mess. But knowing something and proving it are very different things.
We ran user surveys, conducted interviews, and threw everything into a Miro board for affinity mapping. What came back was a clean signal buried under a noisy interface:
Students rated themselves 4–5/5 on Canvas confidence... yet were still missing deadlines
The To-Do list sidebar was the most popular workaround , people were hacking the system
Many students were manually copying due dates into personal planners
Announcements were getting lost in the noise; inconsistent instructor organization was recurring
A classic case of a system that technically works but creates enormous cognitive overhead. Students weren't failing Canvas , Canvas was failing them.
Four requirements. One bot that stole the show.
Based on our research, we focused on four core requirements: Organization, Customization, Calendar Integration, and a Bot.
We iterated from rough sketches → low-fi wireframes → higher-fidelity prototypes. The redesigned dashboard got a proper To-Do list with filters (by class, due date, content type), theme customization, dark mode, and a Google Calendar import button that Canvas has absolutely no business not having already.
The Bot , friendly, zero judgment, lots of answers
Across all courses, sorted and readable
Ones instructors forgot to add to the calendar
So you're not drowning in notifications
One tap to surface what matters
"Hi Tanvi! I'm your helpful FAQ bot. Let me know how I can help you?" , No passive-aggressive reminder emails. No guilt. Just useful.
Designing is fun. Measuring whether it works? That's where it gets interesting.
For CMPM 290, my partner Sonia and I asked: can we quantify the impact of an AI assistant on student assignment completion?
Students with the AI assistant will miss fewer assignments and complete work on time.
The Metric
Running the Numbers
We set a baseline on-time rate of 70% and targeted a 5 percentage point improvement to 75%. Using the standard power analysis formula (α = 0.05, power = 0.80, variance = 0.21):
Rounded up: 1,350 per group, 2,700 total participants , about 9% of a 30,000-student university, rounded to a clean 10% rollout.
The A/B Comparison
| Current Canvas | Proposed Experience |
|---|---|
| Manual navigation | Intelligent, conversational guidance |
| Static calendar due dates | Prioritized task list |
| No detection of missed work | Alerts for unseen assignments |
| No reminders | Personalized, proactive reminders |
| No planning help | Automated or manual work schedule |
We didn't just hype the solution. We stress-tested it.
Would students stop self-regulating entirely?
Could the bot become the very annoyance it was solving?
Academic data is sensitive; AI access to it requires careful governance
Would the algorithm favor certain types of assignments unfairly?
These aren't hypothetical worries. They're the kinds of issues that sink real product launches.
Good UX is both felt and measured.
- ✦The community design project gave us the why: students aren't disorganized, they're navigating a system that buries information and offers no intelligent guidance.
- ✦The online experiment gave us the how much: a 5-point lift in on-time rates, across thousands of students, is worth building for.
- ✦Doing the same problem twice , once through a qualitative, community-centered lens and once through a quantitative, experimental lens , taught me something a single course never could.
- ✦Neither answer is complete without the other. And Canvas really needs a Google Calendar import button. I've now said this in two academic presentations and a blog post. I'm manifesting it.