Code: Three | Devpost

Inspiration

Law enforcement agencies waste up to a third of each shift on manual paperwork from body camera footage. This reduces frontline time, introduces errors, and compromises transparency. With 56% of officers spending three or more hours per shift on documentation, we saw a critical opportunity to enhance efficiency, reduce costs, and allow officers to focus on community engagement and proactive policing. Learning about cases such as the NYPD spending $1.4 million on overtime costs due to additional reporting solidified our mission to create an impactful solution.

What it does

Code: Three is an AI-powered platform designed to streamline documentation for law enforcement by automating report generation from body camera footage. The platform seamlessly integrates with existing camera systems, transcribing real-time audio and video into structured, comprehensive reports. Utilizing advanced Visual-Language Model (VLM) technology, specifically the state-of-the-art (SOTA) CogVLM-v2 model and Whisper API for audio transcription, Code: Three significantly reduces manual paperwork, enhancing operational efficiency. We’re building it to cut administrative burdens and improve efficiency in law enforcement—all driven by firsthand insights from police and our passion for defense tech initiatives.

How we built it

We developed a front-end interface enabling officers to upload body camera videos (will be secured later). Our backend—currently hosted on Replicate—leverages the CogVLM-v2 model to process video footage and Whisper API for accurate audio transcription. We then created custom AI agents to efficiently parse, divide, and analyze video content, returning detailed timelines/transcriptions and identifying key frames essential for thorough reporting.

Challenges we ran into

Our biggest obstacle was handling context for lengthy video footage. CogVLM-v2 struggled to comprehend extensive video data, leading to the loss of critical contextual information. We overcame this by implementing strategic chunking and advanced feature abstraction techniques, maintaining accuracy while keeping essential context intact throughout extensive recordings.

Accomplishments that we're proud of

We have developed a near-SOTA AI tool tailored to the defense/policing sector, setting new standards in surveillance and report-generation technologies.

What we learned

Processing video data efficiently is really f***ing hard.
Maintaining proper context length is crucial for accurate analysis and reporting.
Hosting and managing substantial amounts of data involves a lot of $.

What's next for Code: Three

Our goal is to establish strategic partnerships with simulation companies, local law enforcement agencies, and governmental organizations. Through these collaborations, we aim to deploy Code Three as a premier defense-tech and gov-tech initiative, revolutionizing law enforcement documentation, improving officer efficiency, and promoting community safety.