I make probabilistic systems operationally accountable.

A decision-first reference for reliable tool definitions, parallelism, failure handling, provider adapters, argument validation, and multi-turn agent loops.

A granular account of QLoRA on Llama 3 8B: the failed runs, the measured gains, the serving surprises, and why fine-tuning was not the system bottleneck we thought it was.

A structural investigation of mixed-trust context, indirect injection, partial defenses, privilege separation, and the residual risk that cannot be patched at the model layer.

Why production agents live or die by planning, state, memory, tools, execution, and operational controls around the model.

How small agent errors become shared state, propagate across a pipeline, and get contained with measurable reliability controls.

How to measure chunk-level retrieval, diagnose misses, and catch index or embedding drift before users do.

A claim-level evaluation method that separates retrieval defects, generator overreach, judge error, and production release risk.

Why quality, security, latency, and rollback evidence should travel together before an AI feature reaches production.

A practical decision rule for choosing explicit states, scoped tools, approval gates, and recovery paths.

How to keep malicious or irrelevant instructions inside retrieved content from becoming executable model behavior.