We’re building Samrian because we spent too much time seeing manufacturing companies treat ISO 9001 as a paperwork exercise rather than a quality tool. Most quality managers spend their day feeding a 'compliance machine' manually cross-referencing floor logs against procedures just to survive the next audit.
We wanted to see if we could flip that incentive structure by making verification cheap. If a system can instantly tell you if a shop-floor record matches a work instruction, you can go back to improving the product instead of just fixing the paperwork.
Technically, we hit a wall early on. Standard RAG pipelines (OCR -> Text -> LLM) were useless for manufacturing. Schematics, complex tables, and handwritten notes just became noise when flattened into a text stream. We ended up moving to a multimodal retrieval approach (using ColPali) so the system 'sees' the page layout instead of just reading a bad transcript. We found this recovered about 70% of the accuracy we were losing to OCR.
We're still in the early stages and manufacturing is a notoriously difficult space to move, but I’d love to hear from anyone who has dealt with 'ceremonial' compliance or has thoughts on using vision models for RAG."
We’re building Samrian because we spent too much time seeing manufacturing companies treat ISO 9001 as a paperwork exercise rather than a quality tool. Most quality managers spend their day feeding a 'compliance machine' manually cross-referencing floor logs against procedures just to survive the next audit.
We wanted to see if we could flip that incentive structure by making verification cheap. If a system can instantly tell you if a shop-floor record matches a work instruction, you can go back to improving the product instead of just fixing the paperwork.
Technically, we hit a wall early on. Standard RAG pipelines (OCR -> Text -> LLM) were useless for manufacturing. Schematics, complex tables, and handwritten notes just became noise when flattened into a text stream. We ended up moving to a multimodal retrieval approach (using ColPali) so the system 'sees' the page layout instead of just reading a bad transcript. We found this recovered about 70% of the accuracy we were losing to OCR.
I wrote a more technical deep-dive on that retrieval problem here: http://samrian.com/blog/hardest-part-of-ai-isnt-the-ai
We're still in the early stages and manufacturing is a notoriously difficult space to move, but I’d love to hear from anyone who has dealt with 'ceremonial' compliance or has thoughts on using vision models for RAG."