I evaluate and improve AI and data systems that need to prove they work.

A system under investigation

1. DATA / RETRIEVAL
2. →MODEL / GENERATION
3. →EVALUATION
4. →RELEASE / BEHAVIOUR

At rest you see the whole system. Select a lens to focus on one decision.

Reliability investigation

1. QUERY
2. →RETRIEVAL
3. →CONTEXT
4. →GENERATION
5. →END-TO-END BEHAVIOUR

I look at how retrieval and generation behave before treating the system as ready for production. The Reliability X-Ray comes later; this lens frames the question.

Fit when the demo works but you still cannot say whether a retrieval-backed system is trustworthy in production.

Same workload, explicit constraints

1. SAME WORKLOAD
2. →SAME PROTOCOL
3. →EXPLICIT CONSTRAINTS
4. →ELIGIBILITY
5. →TRADE-OFF OR STOP

I compare candidates under one protocol and hard constraints. If nothing clears the bar, the honest outcome is no recommendation.

Fit when the decision is which model or configuration to run for this workload, not a universal ranking.

Change against a release boundary

1. CHANGE
2. →EVALUATE
3. →COMPARE TO BASELINE
4. →THRESHOLD
5. →PASS / FAIL
6. →RELEASE / BLOCK

I ask whether a prompt, model, retrieval, or tool change crossed an agreed release threshold, not merely whether a metric moved.

Fit when a change is about to ship and you need a clear protect or block framing before users see it.

Locate the originating constraint

1. DATA
2. →PIPELINE
3. →REPRESENTATION
4. →MODEL / RETRIEVAL
5. →DECISION LOGIC
6. →PRODUCTION BEHAVIOUR

I trace the system to find where the bottleneck actually sits (data, pipelines, evaluation, or delivery) before prescribing a product path.

Fit when the problem is broader applied AI or data work, not only a single Inspect, Compare, or Protect decision.