Back to Home

Agentic VQA System

This project implements a Production-Grade Visual Question Answering (VQA) Pipeline. The pipeline is designed to take an image (e.g., a dashcam frame) and a text prompt, process them using an open-source Vision-Language Model (VLM), and return a descriptive answer.

PythonPyTorchHugging Face TransformersBLIP
GitHub

Key Engineering Features

  • Framework: Built with PyTorch, ensuring efficient tensor operations and GPU acceleration where available.
  • Model Library: Integrated with the Hugging Face transformers library to utilize the BlipForQuestionAnswering architecture.
  • Pre-trained Model: Defaults to Salesforce/blip-vqa-base, which is optimized for visual question answering tasks.
  • Processing: Uses BlipProcessor for multimodal input handling (image + text).