Yeah, this one is a bit out there, but it was super fun. I built an AI that actually understands memes. It looks at the image and the text together using PyTorch, TensorFlow, and GPT 4. I used early and late fusion techniques to combine vision and language, plus Tesseract OCR to grab text from the image itself. The model hits over 95% accuracy on meme classification. It’s not your typical corporate project, but it shows how multimodal AI can understand humor, culture, and context. Definitely one of my favorite experiments.