This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend DriveSafe is a privacy-first AI platform that turns hours of dashcam footage into an intelligent, searchable driving history.
A few of my friends drive regularly and use dashcams. Quite often, someone might cross the road unexpectedly, a close call might happen, or another driver might do something unusual, and they need to find that exact moment in the dashcam footage.
The problem is that finding one specific moment in hours of video can take several minutes of manually scrubbing through the recording.
I built DriveSafe to make this simple. Instead of searching through the entire video, AI identifies important moments automatically and helps you quickly find the exact part of the drive you are looking for.
DriveSafe automatically identifies potentially important moments such as pedestrians, cyclists, close vehicle encounters, and sudden scene changes, creates clips and an interactive timeline, and lets you search historical drives using natural language.
🚗 **Live Demo:** [https://drivesafe-gs66.onrender.com/](https://drivesafe-gs66.onrender.com/)
🎥 **Video Demo:** [https://drive.google.com/file/d/1eD3sBaNlxmHwgaiHW-d83Q-_5YtZhKxQ/view?usp=sharing](https://drive.google.com/file/d/1eD3sBaNlxmHwgaiHW-d83Q-_5YtZhKxQ/view?usp=sharing)
💻 **GitHub:** [https://github.com/Mahir-Neema/drivesafe](https://github.com/Mahir-Neema/drivesafe)
DriveSafe is built around Google Gemma 4, with Temporal handling the durable video-processing workflow.
The pipeline is:
Dashcam Video → Temporal Workflow → Frame Extraction → Gemma Vision Analysis → Event Detection → Clip Generation → Embeddings → MongoDB → Searchable Timeline
I use Gemma 4 26B for cloud-based analysis and Gemma through Ollama for local, offline inference. Events are embedded using nomic-embed-text and stored in MongoDB for vector search across historical drives.
The project can also be started locally with Docker, including MongoDB and Ollama.
Dashcam footage can contain highly personal information, so privacy was a major consideration.
Using open-weight Gemma models with Ollama, DriveSafe can process footage completely locally. In Local Edge Mode, video frames and telemetry remain on the user's machine instead of being sent to external cloud servers. Open technologies also made it possible to combine Gemma, Ollama, Temporal, FFmpeg, MongoDB, and Docker into one end-to-end system.
Implements Atlas Vector Search over 768-dimensional embeddings to create long-term historical dashcam memory and enable natural-language retrieval across previous drives.
Orchestrates a resilient, durable VideoAnalysisWorkflow with real-time progress queries, automated retries, heartbeats, and decoupled activity workers.
Uses Gemma 4 vision models in a dual-inference setup:
DriveSafe combines AI-powered video understanding with a complete end-to-end product experience, including 1-click Docker deployment, interactive timeline visualization, FFmpeg clip generation, vector search, historical drive memory, and automated verification tests.