GrabIt

Lobby
Create or join
Playing

Inspiration

Most AI projects today stay in the comfort zone of chat interfaces. I wanted to explore a different direction: AI as an active participant, capable of observing the physical world and making decisions in real time.

The idea for GrabIt emerged after struggling with overly complex concepts. Once the focus shifted to a simple interaction — showing real objects to an AI — the project became both feasible and meaningful.

What it does

GrabIt is a real-time multiplayer game where players compete by showing real-world objects to their webcams. An AI generates live challenges, observes each player’s video stream, understands the intent behind the objects shown, and validates the first correct response.

The AI acts as the game master, referee, and rule enforcer throughout the match.

How we built it

GrabIt was built as a real-time web application:

A web frontend handles video capture and the game interface
A backend manages rooms, players, and real-time synchronization
A multimodal AI (Gemini) analyzes video frames, reasons about objects and concepts, and determines the winner of each round
All video analysis is performed live, with no storage of webcam data.

Challenges we ran into

Some of the main challenges included:

Designing challenges that require conceptual understanding rather than simple object recognition
Handling ambiguity in real-world objects while keeping the game fair
Managing real-time latency and synchronization across multiple players
Balancing AI accuracy with a fast and engaging gameplay experience

Accomplishments that we're proud of

Building a fully playable real-time multiplayer game in a short time frame
Successfully integrating multimodal AI into the core game loop
Creating an AI that acts as an autonomous referee, not just an assistant
Delivering a complete experience from concept to live demo

What we learned

This project highlighted the importance of simplicity in design. A clear interaction model enabled deeper use of AI capabilities than a more complex concept would have. We also learned how powerful multimodal AI can be when it is embedded directly into real-time systems.