01Insights / technical field notes

Technical notes for difficult AI systems.

Field notes connect technical questions to experiments, evaluation surfaces, and engineering decisions. The current archive focuses on evaluation design, RAG regressions, agent failure modes, and release gates; generic examples and evidence diagrams stay explicitly labeled.

Browse the field notes

Diagram showing evaluation evidence branching to ship, change, and stop decisions. — **How to Decide Whether an AI Workflow Ships, Changes, or Stops**Evidence is only useful when it changes the release decision.Diagram by EAVAE Labs.

02Complete index

Read by the question or engineering decision at hand.

Each current note connects a technical pattern to an inspectable artifact, release gate, or engineering action. The archive does not present generic examples as client evidence.

06 notes / reviewed technical guidance

01Reviewed Jul 12, 2026 · 3 sectionsHow to Decide Whether an AI Workflow Ships, Changes, or StopsA decision framework for AI teams that need evidence before shipping, revising, or stopping an agent, RAG, or model workflow.02Reviewed Jul 12, 2026 · 3 sectionsRAG Regression Testing: What to Track Before Retrieval ChangesA concise checklist for tracking RAG regressions across retrieval, context assembly, answer quality, and release gates.03Reviewed Jul 12, 2026 · 3 sectionsFailure Taxonomies for Agent WorkflowsHow to group agent failures by trigger, owner, severity, and release impact so evaluation results lead to engineering action.04Reviewed Jul 12, 2026 · 3 sectionsRelease Gates for LLM FeaturesHow to define blockers, warnings, thresholds, and decision records before releasing LLM-backed features.05Reviewed Jul 12, 2026 · 3 sectionsWhat an Evaluation Plan Should ContainThe minimum useful contents of an AI evaluation plan: target behavior, risks, inputs, replay rows, thresholds, and decision owner.06Reviewed Jul 12, 2026 · 3 sectionsWhy Offline Scores Miss Production AI FailuresWhy offline metrics can miss production AI failures, and how replay rows and failure taxonomies make evaluation more actionable.

09Safe first step

Have an applied AI question these notes do not resolve?

Bring the system boundary, what is uncertain or failing, and what the team needs to learn, build, or decide. Start with sanitized context; private access comes later if the scope requires it.

Discuss a technical problem Inspect the handoff

No credentials, production data, customer records, or private repository access in the first brief.

Prefer to talk it through? Request a 30-minute call