Students lose hours between classes to scrolling, because planning takes longer than the gap itself. Events are scattered across calendars, emails and Instagram. A simple errand can take six apps. And many students skip events rather than walk in alone, since meet-up apps overshare.

The real gap isn't more information. It's timing (what fits right now) and company (someone to go with).

What it does

SC Iris is a campus assistant designed for smart glasses. Our prototype simulates the glasses experience in a web app.

  • Plans your day around your classes, with walk times.
  • Finds what fits: talks and clubs that match your interests and free time.
  • Answers any request: say "I want to buy avocados" and get open stores within 2.5 miles.
  • Offers company, safely: walk into events with other students. Nothing is shared until you approve.
  • Built for USC: students are verified with their @usc.edu email, and events come from USC calendars and club feeds, ranked by source trust (official, club, student).
  • Keeps students engaged: tag-along matching gets students out to events instead of skipping them, and a feedback loop learns from every "go" or "skip" so suggestions improve over time.

A few option cards appear in a simulated lens view, and the student picks one with a voice-style command (go now, go later, tag along, scenic route).

How we built it

  • Simulated glasses interface: a web app that mimics the in-lens display and the voice and gesture picks.
  • Cloud: FastAPI and a LangGraph agent, with Gemini for intent through function calling.
  • Data: Supabase Postgres with PostGIS, pgvector and Realtime, plus Google Maps and Places.
  • USC integration: a usc_verified flag on every profile, USC event and club feeds (seeded demo data in the prototype), and campus-aware walk times between buildings.
  • Engagement: realtime tag-along pings through Supabase Realtime, with a live count of who is going.

Our principle is the AI suggests, code decides. Gemini handles language. Code handles every check.

Stage 1: feasibility (hard filter). An option is kept only if it fits inside the student's free window:

$$ t_{\text{leave}} + t_{\text{walk}} + t_{\text{duration}} + t_{\text{next}} + 5\ \text{min} \le t_{\text{window end}} $$

Stage 2: opportunity score. Options that pass are ranked by:

$$ \text{score} = 0.40\,I + 0.25\,C + 0.15\,S + 0.10\,T - 0.10\,W $$

where $I$ is interest (cosine similarity between the student and event embeddings), $C$ is career fit, $S$ is social signal, $T$ is source trust (official $1.0$, club $0.7$, student $0.4$), and $W$ is walk cost (walk minutes divided by free minutes).

Only the top four options are shown. A learning loop nudges the weights by $\pm 0.02$ based on whether the student chooses "go" or skips.

Consent is a state machine (requested → pending → approved → active → expired, with declined and revoked branches), enforced by Postgres row-level security, so the database itself refuses to reveal a profile without approval.

Challenges we ran into

  • No physical glasses: Meta's display access is in developer preview, so we simulated the lens experience and designed the system so real hardware can plug in later.
  • Tiny display constraints: four options, one line each, even in simulation.
  • Trustworthy AI: an LLM can suggest something that doesn't actually fit between classes, so we moved all feasibility and safety checks into code.
  • Privacy: matching people without exposing profiles before consent.

Accomplishments that we're proud of

  • A working simulated flow from request to a ranked, feasibility-checked set of options.
  • Consent enforced in the database, not just in app logic.
  • Tag-along matching that shares only what the student approves: never location, phone or email.
  • A design ready for real glasses.

What we learned

  • Students need the right option at the right moment, not more information.
  • Designing for a tiny display forces clarity.
  • Splitting LLM and deterministic code makes AI easier to trust and debug.
  • Privacy works best when built into the data layer from day one.

What's next for SC Iris

  • Next: integrate real Meta glasses and run a campus pilot with official USC event feeds, a few clubs and one library floor.
  • Then: deeper engagement through club pages, friends going and reminders, plus hands-free and accessible modes.
  • Later: a walk-with-me safety mode, and reusing the engine on other campuses.

Built With

Share this project:

Updates

Submission history