AI Solutions Locations All Services Website Design Perth SEO Perth Google Ads Perth Facebook Ads Perth Content & Video Real Estate Videography Shopify Perth AI Services Perth Case Studies Free Strategy Call
AI Case Study. Multi-Site Mining Operator

Cross-site parts matching that refuses to be confidently wrong

Matching the same part across two sites' catalogues using evidence tiers, so stock can be shared instead of re-bought.

Client
Multi-site mining operator
Industry
Mining
Scope
Two site catalogues
Anchor
Supplier part number
Tiers
STRONG and CANDIDATE
Evidence Tiers
Site Catalogues
False Match, Now A Rule
Down
Only Way A Tier Moves
01
Before

The problem

Two sites held the same parts under different codes.

Nobody could tell reliably whether site A's item existed at site B. So stock was ordered new while the same part sat idle down the road.

Same part, different codes
Each site's catalogue used its own codes, so the same physical part looked like two different items.
Bought while idle
Stock was re-bought at one site while the other site already held it.
02
The build

What we built

A matching capability inside the agent, built around one anchor: the supplier part number.

A matching number alone is never enough. Corroboration is required before any match is called an equivalent.

The whole design leans one way. When the evidence is thin, the system claims less.

Anchored, then corroborated
The supplier part number anchors the match. Brand, description and supplier family have to back it up before it is called an equivalent.
Claims less, not more
When the evidence is uncertain, the tier only ever moves down.
03
Under the hood

How it works

Every potential match lands in a tier, and the tier is earned.

  • STRONG tier when the part number and the evidence agree across brand, description and supplier family.
  • CANDIDATE tier when the number matches but the evidence does not. A candidate is presented as a lead, never an equivalence.
  • Placeholder part numbers shared by hundreds of unrelated items are screened out first.
  • Every non-match carries the data fill rate, because a missing number never means the other site does not hold the item.
  • When uncertain, the tier only ever moves down.
The placeholder trap
Some part numbers are placeholders shared by hundreds of unrelated items. They are screened out before matching begins.
A non-match is not a no
Every non-match carries the data fill rate. A missing number never means the other site does not hold the item.

During testing, the matcher produced one false match. A single digit had been transposed when a code was composed.

That failure became a permanent rule. Codes are copied character for character from returned rows, and near misses are detected automatically.

The proof here is not a clean run. It is that the one error found was turned into a check that runs every time.

Found in testing
1
A false match caused by one transposed digit, caught before it reached a user.
The rule it left
Char for char
Codes are copied character for character from returned rows, never re-typed.
Near misses
Detected
Codes one character apart are now flagged automatically.
Staff ask “does the other site have this” and get an answer with its evidence, holdings and usage side by side.