ChemicalEngine

1. Need

   A chemical supplier has a price list of 67,102 articles and customers who ask in words other than the catalogue's. The real questions are different: what does a cosmetics quality-control lab buy? Which buffer for this pH? What is the catalogue missing?

2. How it was worked out

   From price list to knowledge base: every substance identified by CAS and not by name, every classification tag shown with the text that produced it, measured always kept apart from inferred. Rules live in tables, not in code: a new rule is a row.

3. Data integration

   PubChem and ChEBI for chemical identity, Rhea reactions (41,672), published procedures (72,608), safety data sheets from the supplier's site; eighteen external services registered, each with its status declared.

4. AI

   Meaning-based search, with local models, finds the product even when the customer's words do not match the catalogue; “paste a list” resolves a whole order. No language model in the search path: the same questions always give the same answers.

5. Solution

   Market baskets by kind of laboratory, guided solutions, twenty related products each with its reason, and the list of catalogue gaps with the number of procedures that justifies each one.

Time saved

An order pasted in any vocabulary resolves itself; the operator checks only what was not recognised.

Knowledge added

What the system does not recognise stays “not found”, never guessed; and the catalogue learns what it is missing.

Without the knowledge matrix, AI does nothing: here it comes last, on top of data put in order.