Inspiration

Bystander-intervention training already exists, but most of it comes in the form of lectures and slide decks that students sit through and forget. Then a difficult moment comes up: a friend is being pressured at a party, someone is being followed in the library, a private photo is circulating in a group chat. People freeze. Most of them care about what is happening, but they have never practiced what to do. Pilots train in simulators and athletes run drills before games, yet students facing the hardest social moments of college get a one-hour video-based training. We built Wick to give them a place to practice.

What it does

Wick lets students rehearse stepping in. A user picks a scenario and has a short voice conversation with AI characters: a person causing harm, an uncomfortable friend, or a passive onlooker. When the session ends, Wick scores the attempt and gives calm, specific coaching on what to try next time.

How we built it

Wick has a React and Vite frontend deployed on Azure Static Web Apps at our custom GoDaddy domain, wick.courses, with a Node.js backend running on Azure Functions. Azure OpenAI runs the role-play engine. Azure AI Speech converts the user's voice to text, and ElevenLabs gives each AI character its own expressive voice. Sessions, streaks, and score data are stored in Azure Cosmos DB.

The home screen is designed as a night map of a grove with trailheads. Each trailhead leads to a realistic campus scenario: The Upstairs Invite (pressure at a house party), The Private Group Chat (non-consensual sharing), and The Library (repeated unwanted attention). During onboarding, an awareness screen explains that the AI characters will say pressuring things on purpose, since that pressure is what the user is practicing against. Each trail also shows its own content note before it starts.

After each session, a debrief scores the attempt and teaches strategies by name from the Five Ds (Direct, Distract, Delegate, Delay, Document). Users can study the same strategies in the in-app Field Guide. A difficulty setting (Gentle, Realistic, Intense) controls how hard the characters fight back, and captions and reduced-motion options were part of the build from the start.

Challenges we ran into

The voice loop was our hardest technical problem. We had to keep each ElevenLabs character voice distinct, keep speech-to-text latency low enough for natural turn-taking, and make sure the text fallback always worked if voice failed.

Writing prompts for an AI that applies realistic social pressure without causing harm took constant tuning. The character has to be uncomfortable enough to practice against while staying within safe limits. That is why we designed the awareness screen and content notes as core screens in the app, with the same care we gave the role-play itself.

We built everything within the 24-hour window with no work allowed before the event, so every Azure resource, API contract, and content handoff had to be ready on schedule. This pushed us into a contract-first workflow. We agreed on the API request and response formats before writing code, which let our frontend and backend developers build in parallel without merge conflicts.

Accomplishments that we're proud of

We shipped a live, playable app at wick.courses. A judge can open it and start a role-play in under a minute with no sign-in or setup. The app includes three complete scenarios, a custom scenario-building studio, AI characters that reply in about two seconds, a working coaching debrief, an interactive Field Guide, and a streak system built around the line "keep your wick lit." All of it uses a consistent dark-forest visual identity with custom illustrated art.

We are proudest of how the app handles its sensitive subject matter. Content notes, intensity controls, and feedback derived from the user's own transcript customize the experience.

What we learned

The hardest part of building with AI characters was setting their boundaries. Writing prompts that keep a character pressuring without letting it become harmful took more iteration than any API integration. We also had to design the surrounding experience (warnings, step-out controls, difficulty levels) so that intense practice still felt safe.

Contract-first collaboration let a designer and a backend engineer build a full-stack app in parallel without blocking each other. For sensitive topics, we found that user trust matters most. Every design decision that told users they were in control did more for the experience than any technical feature.

What's next for Wick

We plan to add trails for workplace harassment, K-12 schools, and community settings outside campus. We also want to build deeper analytics so users can track how their responses improve over weeks of practice. After that, we want to pilot Wick with orientation programs, campus safety offices, and student organizations, since those are the settings where students would most likely use it.