- 🔭 Currently working on Dynamic Ensemble Models
- 🌱 Currently learning FastAPI & AI Deployment
- 🧠 Interested in Transformers, Multimodal Models, and Computer Vision
- 🤖 Passionate about building deep learning models from scratch
- 📫 Reach me at shubhamshukla223311@gmail.com
- Built a multimodal OCR pipeline for early modern Spanish manuscripts
- Used Florence-2 for text detection and segmentation
- Fine-tuned Qwen2-VL (2B) for handwritten line transcription
- Integrated Gemini VLM + Judge LLM to compare candidate outputs
- Applied Sauvola thresholding and skeletonization preprocessing
- Achieved CER 8.92% and BERTScore 0.91
- Implemented core Transformer architecture without high-level libraries
- Implemented:
- Scaled dot-product attention
- Multi-head attention
- Positional encoding
- Layer normalization
- Trained custom transformer on sequence modeling tasks
- Built an image classification model using ResNet50 transfer learning
- Applied data augmentation and fine-tuning
- Achieved ~92% validation accuracy
- Developed an end-to-end ML web application
- Built using scikit-learn models
- Implemented real-time prediction interface
- Achieved ~89% test accuracy
💡 I enjoy building AI architectures from scratch to deeply understand how models work.