Writing

Notes from building retrieval systems that have to survive an audit.

Engineering notes by , Sr. AI/ML Engineer in Atlanta, GA. Each of these started as a short note on LinkedIn — a diagram, a rule of thumb, a list of five things that work. Here they're written out properly: the reasoning underneath, the steps to implement it, and code you can run.

AI-Assisted CodingArchitectureChunkingCode QualityDeveloper ToolingEvaluationGraph-RAGGuardrailsKnowledge GraphsLLMMLOpsPrompt EngineeringPythonRAGRetrievalShipping
  1. 019 min readEvery stage of a RAG pipeline needs its own evaluation gateA single end-to-end score says the system got worse but never says which stage broke; per-stage gates turn a vague regression into a located one.RAGEvaluationMLOps
  2. 029 min readGraph-RAG needs eight evaluation gates, not oneGraph-RAG fails in four places a vector index cannot, so extraction, graph construction, community detection, and summarisation each need their own measured gate.Graph-RAGEvaluationKnowledge Graphs
  3. 039 min readPrompt chains: deterministic pipelines that beat premature agentsFixing control flow at author time is what separates a prompt chain from an agent, and it buys predictable cost, reproducible traces, and testable steps.LLMPrompt EngineeringArchitecture
  4. 0411 min readSizing chunks so Graph-RAG extraction never loses an edgeA four-tier routing rule that picks chunk size from document length, so every unit handed to the extractor can still yield at least one complete triple.Graph-RAGChunkingRetrieval
  5. 059 min readFive layers between a retrieval system and a fabricated answerHallucination is several distinct failures under one name; five composable controls — grounding, instruction, verification, evaluation, guardrails — address them.LLMGuardrailsEvaluation
  6. 069 min readCrossing the gap from Python script to double-clickable appA v1.0.0 release maps the three rungs between a working script and something a non-developer can open, and what each rung costs the person who built it.PythonShippingDeveloper Tooling
  7. 079 min readPre-commit hooks are the feedback loop AI coding was missingAn assistant that cannot see its own lint errors depends on a human to relay them; deterministic hooks turn that relay into a signal the machine can read.Developer ToolingAI-Assisted CodingCode Quality