eliteAI / Evidence / How it works
How the three models think
Planning
Accuracy on the demo data
91 of 100 seeded duplicate pairs resolved correctly at the default threshold.
The question it answers
Two people wrote two different sentences. Are they asking for the same part?
What it reads
- request, the free text a department typed
- catalogue, the official item code
- raisedBy, the department
- equipment, the tag the part sits on
- leadTimeWeeks
The technique
Text matching and entity resolution. Requests are normalised, Indonesian and English terms are mapped onto one vocabulary, then token overlap and the equipment tag decide whether two requests resolve to the same catalogue record.
Where a language model is used
A language model is used here, and only here, to suggest synonym pairs such as pompa and pump. A person approves each pair before it enters the vocabulary.
The match itself is not written by a language model. It is token overlap and a threshold you can move on the Planning screen.
What it cannot do
It cannot tell you whether the two departments need the part at the same time, so one merged order can still be the wrong answer.
Spare parts
Accuracy on the demo data
86 percent of suggestions accepted without change on the demo data.
The question it answers
Out of 1,240 items, which handful is worth a decision this month, and should the level go up or down?
What it reads
- onHand and suggested
- 24 months of movement history
- unitIdr, the value per unit
- equipment and category
- supplier lead time, including changes to it
- insurance, the criticality flag
The technique
An intermittent demand forecast with gradient boosted trees. Most of these parts move a few times a year, so the model predicts the chance of demand in a period and the size of it separately, then compares cover against lead time.
Where a language model is used
No language model is involved in this module at any point.
Nothing here is generated text. Every number on the screen comes from the movement history and the lead time record.
What it cannot do
It cannot see a criticality judgement an engineer holds in their head, which is why insurance stock is protected by hand and never proposed for cutting.
Finance
Accuracy on the demo data
97 percent of the 25 seeded invoices classified correctly line by line.
The question it answers
Does this invoice line follow the rate the contract says it should?
What it reads
- invoiceNo, supplier and contractNo
- each line, its qty and invoiced rate
- the contract rateCard, its agreed rate and clause number
- the date the contract was signed
The technique
Document extraction to pull the lines off the invoice, then a rules engine that compares each line against the clause in the rate card and computes the arithmetic of the difference.
Where a language model is used
A language model is used for extraction only, to read a line off a scanned invoice. It never decides whether a line is an exception.
The comparison is a rule against a signed clause. The same input always gives the same answer, and the clause is shown next to it.
What it cannot do
It cannot see a variation agreed in a side letter that was never loaded into the contract record, so a correct rate can still be flagged.
Three separate models, not one chatbot. Two of them never call a language model at all.