When the evaluation says pass, but the agent is wrong.
Deterministic checks, semantic judges and the difference between a score and a guarantee.
Explore the topic ↗Why do systems fail in unexpected ways? What makes them easier to operate? Explore the investigations, experiments and tools behind our engineering work.
NOTRIX / SOFTWARE SYSTEMS · APPLIED AI · RELIABILITY
Failure modes, overlooked tradeoffs and experiments that reveal how software actually behaves.
The quiet degradation of retrieval quality — and what normal monitoring misses.
Deterministic checks, semantic judges and the difference between a score and a guarantee.
Explore the topic ↗From proposed actions to guarded execution, traces and completion decisions.
Visit the labs ↗Systems thinking, tradeoffs, architecture notes and practical engineering judgment.
Browse engineering notes ↗Explore bounded demos of agent behavior, tool calls, evaluation and system traces.
Inspect planning, execution, evaluation and failure recovery in approved demo scenarios.
Tools that make specific, recurring tasks easier — not technology for its own sake.
Capture and compare AI workflow execution to understand what changed, where and why.
Turn working material into publishable drafts, without recreating the workflow from scratch.
Focused publishing and experimental spaces, connected by engineering questions rather than corporate categories.
Architecture, software engineering and practical systems thinking.
Failure analysis, observability and evaluation of AI workflows.
Interactive systems, bounded experiments and agent traces.
We also work directly with teams on persistent engineering problems — with practical diagnosis, clearer tradeoffs and focused implementation.
Start a technical conversation ↗Scoped technical work. No generalized “digital transformation” promises.