Hackathon CoPilot

Hackathons are engines of innovation, but they suffer from a severe "Mentorship Gap." Organizers cannot physically mentor hundreds of teams simultaneously, leading to technical blockers and participant burnout. Hackathon Copilot solves this by digitizing the organizer. It is a multimodal "Hackathon Operating System" that uses Gemini 3 Pro and Veo to create an always-available, intelligent "Founder Twin" that can mentor, debug code via screen sharing, critique video pitches, and write scripts—scaling one human leader to infinite participants. 🛠️ Gemini 3 Integration (Technical Execution) Our application is not a wrapper; it is a complex orchestration of the entire Gemini 3 ecosystem. We utilize specific models for distinct cognitive and sensory tasks: Gemini 3 Pro (Reasoning & RAG): The core "Brain." We leverage the massive context window to ingest the organizer's specific "Knowledge Base" (PDF rulebooks, API documentation, and codebases). This allows the AI to answer context-aware questions with citations, ensuring 100% adherence to hackathon rules. Gemini 3 Vision (Visual Debugging): We implemented a "Visual Stack Trace" feature. Participants can upload screenshots or stream video of their terminal errors. The model performs pixel-level analysis to extract error logs and suggest code fixes, bypassing the need for copy-pasting text. Gemini Live API (Real-Time Multimodal): This is our technical crown jewel. We implemented the Live API (gemini-2.5-flash-native-audio) via WebSockets to enable sub-500ms voice interactions. We built a custom pipeline that streams raw PCM audio and video frames (Screen Share) to the model, allowing the AI to "watch" a user code in real-time and offer verbal guidance. Veo 3.1 (Generative Avatar): To humanize the AI, we use veo-3.1-fast-generate-preview. When the "Founder" gives advice, the app generates a lifelike video of the organizer's avatar speaking the text, creating a high-fidelity "Digital Twin" experience. Google Search Tooling: For pitch analysis, if a user submits a YouTube link, we utilize the googleSearch tool to reconstruct the video's context (metadata, transcripts) to provide a fair critique even without raw file access. ✨ Key Features (Innovation & Wow Factor) The Live Uplink: A real-time video call with the AI Founder. Participants can click "Share Screen," and the AI will analyze their code editor live, identifying bugs and architectural flaws instantly. Dual-Mode Pitch Simulator: File Mode: Uses Gemini Vision to analyze slide density, body language, and visual storytelling frame-by-frame. Link Mode: Uses Search Grounding to grade content hosted on external platforms like YouTube. Neural Scriptwriter: A specialized agent that reads the user's project README.md and synthesizes a perfectly timed 2-minute pitch script, optimized for the specific judging criteria of the event. Founder Control Deck: A dashboard for organizers to "upload their brain" (documents) and "upload their face" (Veo avatar), deploying a custom instance of the OS in seconds. 🌍 Potential Impact Hackathon Copilot democratizes access to high-quality mentorship. For Organizers: It removes the bottleneck of answering repetitive questions ("What's the wifi?", "Is this eligible?"), freeing them to focus on high-value logistics. For Participants: It provides a 24/7 Senior Engineer and Pitch Coach, leveling the playing field for beginners who might otherwise drop out due to technical blockers. Scalability: While built for hackathons, this architecture is immediately applicable to Education (TA bots) and Enterprise Onboarding (HR/Tech Support twins). 🏗️ Architectural Diagram & Stack Frontend: React 19, TypeScript, Vite, Tailwind CSS. AI Engine: @google/genai SDK. Audio Processing: Custom AudioContext nodes for PCM16 conversion (Browser ↔ Gemini). State Management: IndexedDB for local persistence of large media assets (ensuring the "Founder" stays alive across reloads).

Built With

• aistudio
• gemini3
• typescript
• veo-video-converter