Skip to content

AI Systems

The architect's job - components, trade-offs, NFRs, teardowns - applied to LLM-powered systems. Hands-on notes, not rehashed blog posts.

  • rag/ - retrieval-augmented generation: chunking, embeddings, vector stores, hybrid search, reranking, evaluation.
  • agents/ - LangGraph & orchestration: state machines, tool use, memory, multi-agent, human-in-the-loop.
  • extraction/ - intelligent/structured extraction from unstructured docs: schemas, confidence, validation.
  • agentic-data/ - agentic data intelligence: text-to-SQL, semantic layers, query planning over warehouses.
  • rules/ - natural-language to executable business rules: DSLs, verification, determinism vs. LLM judgment.
  • llmops/ - Opik & friends: tracing, eval runs, guardrails, cost/latency, prompt management.
  • patterns/ - AI architecture patterns: RAG variants, agent loops, router, reflection, ReAct, orchestrator-worker.

Ripples into the rest of the notebook: AI-native NFRs (eval, hallucination/grounding, token cost, latency, prompt-injection) live in ../cross-cutting; the recurring arguments (RAG vs fine-tune vs long-context, agent vs pipeline, build vs API) live in ../decisions.