When two businesses combine, the savings are sitting in plain sight: the same suppliers, the same parts,
bought twice at different prices. Nobody can see them because the data doesn't line up — two part-numbering schemes,
two supplier masters, two ways of writing the same company's name.
This is that problem, on two small sample catalogues. Press the button and watch it resolve.
Nothing has run yet — the tables below are the raw inputs.
Step 1 · The inputs
Two files that don't agree on anything
Different column names, different part numbers for the same item, the same supplier written three ways,
mixed units and currencies. This is what real extracts look like.
ridgeline_drive_systems_spend.csv
Part no.
Description
Vendor
Cost
HALDEN_OFFHWY_PARTS_EXPORT.XLSX
Item ID
Item description
Supplier name
Unit price
Step 2 · Intake
Parsed, validated, normalised
Columns are found by name rather than position, so a reordered or renamed export still lands.
Anything that fails validation is quarantined with a reason and sent back to the source — never silently dropped.
Exception queue
Record
Rule
What happens
Step 3 · Supplier resolution
Who is actually the same company?
Name variants, legal-entity suffixes and site qualifiers are normalised, then compared against a
registration identifier where one exists. Every merge carries its evidence — and the near-miss below is deliberately
not merged, because precision is what makes a finding actionable.
Step 4 · Part equivalence
Which parts are the same thing?
Attributes are extracted from free-text descriptions — thread, length, strength class, coating,
bearing dimensions — and compared alongside the manufacturer's own part number. A match is only offered when the
evidence supports it, and the reasoning is shown so a category manager can overrule it.
Equivalent is not the same as approved
A match says two part numbers describe the same thing. It does not say
either one may be used. Qualification is a hard filter applied before cost is looked at, not a factor weighted
inside the score — approved source status, part approval (PPAP), customer-directed sources and export
control either pass, or the part is not a candidate at all. A confidence of 0.92 is something for a category
manager to review; it is never an instruction to dual-source.
Step 5 · The finding
The same part, two prices
Costs are shown as an index — the cheaper unit is 100 — which is exactly how we handle a real pilot:
the client sets the base and keeps it, so no actual price ever leaves their control. Variance analysis works
identically either way.
Northbridge Industrial Calder Manufacturing
Equivalent item
Supplier
Indexed unit cost
Gap
Step 6 · With your data
What we'd actually need — and what stays with you
Comes to us, in the clear
Supplier name, address, registration ID
Duplicate detection compares name and address variants. Hashed names can't be fuzzy-matched, so masking them would defeat the exercise.
Part number, description, specification, UOM
This is the matching evidence. It already appears on drawings and in supplier catalogues.
Never leaves you
Unit cost
Indexed to a base you set and keep. We can find the variance without ever seeing the price.
Volumes
Banded, so findings can be weighted without disclosing production rates.
Contracts, rebates, payment terms
Excluded — not needed for overlap detection.
Controlled & personal data
Excluded at extract time. We don't want it and aren't equipped to hold it.
What this demo is, precisely
Fictional companies, fictional part numbers, fictional costs. The matching you just watched is
computed in your browser from the two sample tables — it isn't a recording, and there is no server. What it
demonstrates is the intake and normalisation layer: ingest, validation, supplier resolution, part equivalence and
variance. Sourcing decisions stay with your buyers; nothing here selects a vendor or commits spend.
That was sample data. The real thing runs on your own supplier spreadsheets —
uploaded, checked and turned into quotable pricing.