OmniVision AI

• 

  Receipt AI (Overview Dashboard)

• 

  Extract (Receipt photo upload / text paste section)

• 

  Model Performance (Model comparisons: Fine-tuned flan-t5-small vs NVIDIA Nemotron)

• 

  Guide (Documentation & dashboard instructions)

• 

  Overview ("How it works" & metrics overview)

Inspiration

In today's fast-paced environment, processing and making sense of high-volume visual data in real time remains a significant computational challenge. We were inspired to build a solution that leverages modern cloud GPU capabilities to deliver low-latency visual analytics directly to users.

What it does

OmniVision AI is an intelligent visual processing platform designed to analyze video streams and images in real time. It extracts automated insights, identifies key metrics, and provides an intuitive, interactive dashboard for live monitoring and evaluation.

How we built it

• Infrastructure & Acceleration: Deployed on Nebius Cloud using NVIDIA GPUs for high-throughput, low-latency machine learning inference.
• Backend Services: Developed using Python and FastAPI to handle asynchronous stream processing and model deployment efficiently.
• Frontend & Visualization: Built with React to provide real-time status updates and visual metrics.

Challenges we ran into

• Optimizing GPU memory allocation and frame ingestion pipelines to prevent bottlenecks during multi-stream processing.
• Minimizing visual inference latency to deliver accurate, sub-second feedback on live data.

Accomplishments that we're proud of

• Successfully created an end-to-end computer vision pipeline optimized for Nebius Cloud GPU infrastructure.
• Delivered a sleek, real-time dashboard capable of processing high-resolution visual inputs.

What we learned

• Best practices for structuring scalable, concurrent computer vision pipelines.
• Leveraging specialized Nebius Cloud hardware configurations for deep learning model inference.

What's next for OmniVision AI

• Expanding multi-camera and live RTSP stream support.
• Integrating fine-tuned domain-specific vision-language models for granular semantic video analysis.

Built With

• computervision
• cuda
• fastapi
• nebius-cloud
• opencv
• python
• pytorch
• react