Why Traditional RAG Fails at Multi-Hop Reasoning (And My Journey Building a GraphRAG Fix)I recently set out with a specific goal: I wanted to build an AI agent that could truly understand a massive, undocumented codebase. If you’ve ever tried to feed a codebase into a standard Retrieval-A18 min read·shalem-raju.hashnode.devMar 22
The Final Assembly: Building the Complete Transformer from Scratch in PyTorchWe finally made it. If you have been following this "learning in public" series from the beginning, we have been deep in the trenches of PyTorch tensor shapes. We’ve built Supermarket Embeddings, inje4 min read·shalem-raju.hashnode.devMar 7
Putting the Pieces Together: Building the Transformer Decoder Block in PyTorchIf you’ve been following my "learning in public" series, we’ve spent the last few posts in the trenches. We’ve wrestled with tensor shapes, built mathematical blindfolds (Masked Self-Attention), bridg4 min read·shalem-raju.hashnode.devMar 7
Decoding the Decoder: Masked Self-Attention and Cross-Attention in PyTorchIf you’ve been following this "learning in public" PyTorch series, we have successfully built the entire Transformer Encoder. We gave it a sentence, mapped it to embeddings, added spatial awareness, a5 min read·shalem-raju.hashnode.devMar 6
Completing the Transformer Encoder: Add & Norm and the Feed-Forward LayerIf you’ve been following my PyTorch "learning in public" series, we’ve already tackled the hardest part of the Transformer architecture: Multi-Head Attention (MHA). We took our tokens, split them into7 min read·shalem-raju.hashnode.devMar 5
