DayWise

Inspiration

Your calendar knows where you're supposed to be. Your productivity app knows what you're supposed to do. But neither knows how you actually work.

As college students, we noticed that traditional planners help organize our schedules but rarely adapt to what actually happens throughout the day. Unexpected free time goes unused, study plans change, and finding the right environment often comes down to guesswork.

We wanted to build an AI agent that doesn't just organize your time, but understands your habits, recognizes opportunities, and helps you stay on track without constant manual planning.

What It Does

We built DayWise, a context-aware AI productivity agent that operates directly through iMessage.

Our agent combines several capabilities:

Proactive planning: Identifies free time and uses opted-in location context to suggest study sessions, with user approval.
Intelligent scheduling: Helps organize calendar events, split study blocks, and propose schedule changes for review.
Workspace analysis: Uses photos and environmental context to estimate lighting, clutter, ergonomics, productivity, and mood.
Personalized recommendations: Learns from previous study sessions to suggest environments suited to the user.
Continuous improvement: Connects productivity, mood, and distraction ratings to completed sessions so recommendations can adapt over time.

Rather than requiring students to manually track every hour, our agent uses available context and lightweight check-ins to make planning easier.

How We Built It

We integrated Photon Spectrum for iMessage communication, xAI/Gemini for conversational intelligence, Google Calendar for scheduling, and Photon/Find My for permission-based location awareness.

Our TypeScript bot coordinates these services, while a separate FastAPI backend supports advanced scheduling experiments and simulated planning workflows. We also use local persistence with optional Supabase synchronization.

For personalized recommendations, we implemented a LinUCB contextual bandit with ridge-based learning, balancing previously successful study environments with promising unexplored alternatives.

Its recommendation score follows:

$$ \operatorname{UCB}(a,x)=x^\top\hat{\theta}_a+\alpha\sqrt{x^\top A_a^{-1}x} $$

This allows our system to consider both predicted outcomes and uncertainty when ranking study spaces.

Challenges We Faced

Our biggest challenge was making several AI capabilities function as one cohesive agent, rather than disconnected features.

We also had to balance proactive automation with user control, handle unreliable location signals and calendar integration, and develop personalized predictions despite limited initial user data.

We addressed these through approval-based scheduling, context checks, synthetic starter examples, persistent session feedback, and distraction-aware learning.

What We Learned

We learned that building an effective AI agent requires more than an LLM. It requires connecting context, reasoning, actions, and feedback into a system that meaningfully reduces user effort.

We also learned the importance of keeping users in control while allowing AI to handle repetitive planning and analysis.

What's Next

We plan to incorporate travel distance and estimated walking time into study-space recommendations, making suggestions more practical based on the user's current location and available time.

We also want to expand contextual activity understanding, improve adaptive calendar rescheduling, and learn more about individual productivity patterns across different times and tasks.

Our vision is simple: traditional planners organize what you intend to do. We're building an agent that learns from what you actually do to help make your next day better.