Guides
Engineering guides
Guides written from production systems: LangGraph architecture, MCP OAuth, RAG evaluation with RAGAS and LLM judges, GraphRAG vs vector RAG, caching, routing.
LangGraph production architecture
LangGraph production architecture: what I learned building an eight-agent system
How a production LangGraph system is built: a supervisor over specialized agents, typed state, checkpointed steps, bounded tool calls, evaluation as the gate.
MCP OAuth architecture
Production MCP OAuth architecture: connectors, relays and the security mistakes to avoid
OAuth for MCP servers in production: PKCE, token audience, a relay that keeps provider tokens off agents, per-tool consent, a call ledger and SSRF defenses.
RAG evaluation
RAG evaluation with RAGAS and a claim-level LLM judge (and why model choice was not the bottleneck)
How I evaluate RAG in production: a golden set, retrieval metrics, RAGAS, a claim-level LLM judge for faithfulness and grounding filters, all wired into tests.
GraphRAG vs vector RAG
GraphRAG vs vector RAG: when a knowledge graph earns its cost
When GraphRAG beats vector RAG: multi-hop, aggregate and entity-centric questions, hybrid retrieval with Neo4j, reranking, evaluation and running cost.
LLM cost optimization
Semantic caching, model routing and token-level cost accounting in a multi-provider AI platform
Why LLM costs explode in production, and the four controls that fix it: token-level cost accounting, semantic caching, complexity-based routing and quotas.