C-Drive | Devpost

Inspiration

Our original goal was to build an agentic Claude chatbot that could automate complex tasks using a multi-agent architecture. The backend was powered by FastAPI and LangGraph, while the frontend was a Next.js interface that visualized real-time agent actions through screenshots and progress updates.

Halfway through the hackathon, we realized we wanted to pivot toward something with more real-world impact. After no sleeping from KnightHacks of brainstorming, we repurposed the structure of our project into a driver state detection system capable of identifying signs of drowsiness or distraction. The irony of detecting sleepy drivers after getting no sleep ourselves wasn’t lost on us.

What We Learned

We learned how to move quickly between ideas and re-use existing infrastructure efficiently. Working under pressure, we saw how a robust backend and modular design can make pivoting possible without starting from scratch. We also deepened our understanding of computer vision, model fine-tuning, and how to use Google Colab to train and deploy small but effective models on limited hardware.

How We Built It

We started with a FastAPI backend for multi-agent automation and a Next.js + Tailwind frontend for visualization. Once we pivoted, we built a driver detection module using OpenCV and MediaPipe.

We sourced a labeled dataset from Kaggle, preprocessed the images, and trained a lightweight YOLOv8-nano model on Google Colab using GPU acceleration. The model learned to detect eyes, head pose, and gaze direction. We integrated it with our camera input pipeline, allowing it to run in real time and flag potential drowsiness or inattentive states.

Challenges

Time was the biggest challenge. With very little time to pivot, we had to scale down the chatbot’s complexity and repurpose our codebase to support video input and model inference. Managing dependencies between the AI backend and the real-time detection logic also required several quick iterations. Luckily I came off with training vision at KnightHacks with my robot arm chess player where I used CV for board state detection so finding a dataset and training using YOLO was not too bad.