Engineering notes

Giving AI a narrow job in DeskLab

Interpret flexible language, then enforce product rules through ordinary application code.

DeskLab uses natural language to make a workspace configurator easier to approach. Someone can describe their budget, available width and style instead of filling every field manually. The model's job ends at preferences: it supplies structured values that application code validates before planning a setup. It does not supply authoritative product IDs, prices or purchase actions.

A response schema helps define the expected shape. Business correctness still needs its own checks. OWASP distinguishes syntactic validation from semantic validation, including relationships between values.

In DeskLab, the deterministic planner checks the fictional catalog's availability, total cost and combined footprints. A 100 cm desk still has only 100 cm of tabletop when placed inside a 140 cm space. Accessories fitting individually does not establish that they fit together.

Manual swaps pass through the same validator as generated recommendations. This gives both interaction paths one contract and one place to fix a constraint.

If live interpretation times out or returns unusable output, a labelled rules fallback keeps the preference flow available. That fallback can still misread intent, so preferences remain visible and editable. The boundary limits what interpretation can control while keeping the product decisions inspectable.

Explore DeskLab's AI decision and footprint checks

All engineering notes