4 writing
Writing
Notes on AI systems, retrieval, agents, and what I am learning.
- Building Voice Agents: Cascaded Pipelines, Speech-To-Speech, and the War on Latency
A voice agent is a chatbot where every 100 ms is visible. How the cascaded STT → LLM → TTS architecture works, how it compares to native speech-to-speech models, and the streaming and turn-taking tricks that make it feel human.
- LightRAG: Graph RAG Without the Graph-Sized Bill
GraphRAG is powerful but slow and expensive to build and update. LightRAG keeps the knowledge graph, drops the community summaries, and retrieves at two levels — specific entities and broad themes — at a fraction of the cost.
- GraphRAG: When Your Answers Live Between the Documents
Vector search finds passages that look like your question. It cannot answer "what are the main themes across all of these?". GraphRAG builds a knowledge graph and community summaries so it can.
- RAG, Properly: From Naive to Advanced to Modular
Everyone has built the ten-line RAG demo. This is the guide to everything after it — pre-retrieval, retrieval and post-retrieval techniques, and how the pieces snap together into a modular system.