Master Data (MDM)
One record per real thing,
however many systems say so.
The same company, hotel or person lives in five systems under five spellings. Quadrazene matches them, clusters the duplicates, builds a single golden record by rules you can read, and hands the hard cases to a model and then to a human steward.
- name · best-formatted · CRM
- address · most-complete · ERP
- phone · mode · 3 of 4 agree
Matching
Four tiers, cheapest signal first.
Records are grouped into blocking buckets so only plausible pairs are scored. The first tier that matches decides the score.
Deterministic
A shared identifier, such as a code, tax id or email, in any of four key columns.
Geo + name
Within a set distance and similar enough by name, so different spellings of one place still meet.
Embedding
Cosine similarity between record embeddings, for the pairs no rule would catch.
Phone
Normalized phone equality, whatever the formatting.
Thresholds are configuration, not code. Clusters use union-find, so if A matches B and B matches C, all three land together even though A and C were never compared directly.
Survivorship
Which value wins, written down.
Per-field rules decide what the golden record says, and a log records why. A steward can override any field, and the override sticks.
AI where it earns its keep
Models for the hard pairs, humans for the last call.
LLM adjudication
For a candidate duplicate, the model reads both records, their field-level signals and your domain guidance, then returns merge, dismiss or uncertain, with a confidence.
Auto-act above a threshold
Set the confidence at which the model acts on its own. Anything below it waits for a steward, and auto-acted pairs are labelled as such.
Re-assess without re-paying
Change your decision rules and re-apply them to the responses already stored, with no new model calls.
Hard rules in the prompt
Guardrails such as 'co-located properties under different brands are separate' are part of the domain's guidance, not something a reviewer has to remember.
Bring your model
Adjudicate with a hosted model or a local one, and choose per run. Embeddings use the tenant's existing model connection.
Semantic search
Look up a golden record by text, name or exact key. Approximate-nearest-neighbour search when pgvector is present, with an in-process fallback when it isn't.
Built to run at scale
Your data stays in your database.
External MDM store
Source records, match pairs, clusters, golden records and vectors live in a database you designate, separate from the operational one, with its own retention and access rules.
Duplicate scans as jobs
Full duplicate scans run as background worker jobs with progress, so a large domain does not tie up a browser tab.
Presets, then your fields
Start from a built-in hotel, company or person preset and override field by field. There is no entity-specific code.
Incremental ingest
Load a new extract and re-match without starting over. Existing pairs are kept, not re-scored.
Code-integrity checks
Detect conflicting identifiers, such as one code claimed by unrelated records, and surface them for review.
Export the pairs
Export candidate pairs by bucket, in the columns and format you choose, across the full result set, not just the page on screen.
Part of the platform, not a silo
Master data that the rest of Quadrazene can use.
Ask in the Reactor
Point the Reactor at the MDM System and ask 'which golden records have confidence below 0.5?' in plain English.
Watch it on Radar
Pipeline funnel, cluster quality, daily ingest and match-run history as live dashboard tiles.
Review through the Inbox
A Chain can raise an Inbox item for any low-scoring cluster and route it to the data-steward team.
Govern with policy
A data-quality policy can open a finding when a golden record's confidence drops below a floor.
Home widget
An MDM operations widget puts pending review and recent runs on your Rialto canvas.
Audited
Match runs are recorded, and steward decisions land in the hash-chained audit ledger.
Resolve your own duplicates.
Bring an extract from two systems and we'll show the clusters, the golden records and the pairs the model flagged.