LLMs, agents and the tooling that puts them to work.
Practical strategies for truncating, summarizing, and structuring conversation history to keep LLM responses coherent across hundreds of turns.
Compare function calling and structured output modes in LLM APIs. Understand when each fits, the trade-offs in reliability, latency, and schema complexity.
Implementing server-sent events for LLM streaming, handling backpressure in Node.js, and parsing partial JSON tokens without corrupting output.
A framework for choosing between pgvector, Qdrant, Pinecone, and Weaviate based on query type, filter support, indexing latency, and hybrid search capabilities.
Practical chunking strategies for RAG that survive tables, code blocks, PDF layouts, multi-column text, and other real-world document messiness without losing context.
Token estimates mislead cost and latency. Learn what to measure instead: input/output split, cache hits, tool calls, and retries.
A practical guide to quantization levels for running LLMs on consumer hardware: perplexity, memory, and inference speed trade-offs across GPTQ, GGUF, and AWQ.
Techniques for assessing LLM response quality when no reference answers exist: self-consistency checks, LLM-as-judge, task-specific metrics, and human evaluation proxies.
Practical patterns for building agent loops that make tool calls, handle retries, and survive the failure modes that documentation glosses over.
