Recursive, Kuken Modernize 70-Year Drawing Search

This is practical enterprise AI at its best: a focused retrieval workflow tied to an expensive bottleneck, with engineers retaining final authority. The results are promising, but remain trial metrics rather than proof of autonomous engineering accuracy.

• –Multimodal inputs—including screenshots, PDFs, drawings, and parts lists—make the system useful in real field workflows.
• –Structured extraction matters because engineering archives encode meaning in layouts, revision tables, dimensions, and document relationships—not just text.
• –The drawing-checking feature appears best suited to triage: roughly 90% of flagged prototype issues were genuine, but engineers still validate the results.
• –Faster archive access also addresses knowledge transfer, giving newer engineers more independence while freeing veterans to mentor and handle complex work.
• –The broader developer lesson is that domain-specific schemas, provenance, and workflow integration can matter more than adding a generic chatbot interface.