Making an audit assistant show its work
Written for engineers and plant staff who would check every claim. I designed and built it for a manual lookup process that costs roughly $2.5 million a year across 140 North American plants. It is now licensed for evaluation, with two of nine modules working, and was demoed to the client's packaging forum and taken to its executive steering committee in June 2026.
A packaging audit is a lookup problem wearing a compliance costume. Someone on a plant floor has a question, the answer exists in a governing standard, and finding it takes long enough that people stop looking and start guessing. Multiply that across 140 plants and the guessing becomes the expensive part.
The obvious move is to put a language model in front of the standard. The obvious move is also how you get a system that is confidently wrong in a setting where being confidently wrong has consequences.
Retrieval is the product. The model is the part that phrases things. If retrieval surfaces the wrong section, no amount of prompt work saves the answer, it just makes it read better. Chunking follows the document's own logic. Standards are written in numbered, nested clauses. Splitting on a fixed token count cuts a requirement away from the exception that qualifies it, and the exception is usually the part that matters. "I don't know" is a valid output. A system that always answers is a system that sometimes invents.