Plain-English deep dives on the projects Vrund Patel ships: what was built, the tech behind it, and why.
Why embeddings need their own database — ANN indexes like HNSW and IVF, hybrid search, reranking, quantization, and how to pick between pgvector, Pinecone, Qdrant, Weaviate and Milvus.
· 24 min read · Vector, Embeddings, RAG, Retrieval, HNSW, pgvector, LangChain, LangGraph
What actually happens between hitting enter and reading a reply — tokenization, prefill, causal attention, the KV cache, logits, and why the first token is always the slowest.
· 19 min read · LLM, Transformers, Attention, KV Cache, Inference, Tokenization, MoE
How an agentic RAG pipeline answers questions about a WhatsApp group chat written in English and romanized Gujarati — chunking, glossing, hybrid retrieval, and the choices behind each.
· 16 min read · Agentic AI, RAG, LangGraph, ChromaDB, BM25, Whisper
Resume Roaster & Rebuilder - A plain-English deep dive into the three ideas powering my resume agent — Agentic AI, LangGraph, and the Model Context Protocol — and why I chose each one.
· 7 min read · Agentic AI, LangGraph, MCP, FastMCP, Python