- 4,000+
- Dealer feeds consolidated
- 18
- Supplier pipelines
- 4
- Stages in the resolution cascade
Context
A firearms e-commerce data platform aggregates listings from more than 4,000 dealer feeds arriving through 18 supplier pipelines. The same product attribute turned up spelled dozens of ways, so votes that should have reinforced each other split instead, and the consolidated catalog carried low-confidence, conflicting, and sparsely populated values.
Constraints
A forensic audit of the data put the cause somewhere other than where it was being looked for: there was no semantic layer, so sources were voting on strings rather than on meanings. One manufacturer appeared under 18 different names on a single product, splitting a consensus no resolution algorithm could recover afterwards. Generic matching would have made it worse — merging distinct cartridge families, or conflating an operating mechanism with a trigger mechanism, quietly and at scale.
RegainFlow's role
RegainFlow ran the data audit and architected and built the canonicalization layer — the vocabularies, the cross-reference engine, the resolution pipeline, and the evaluation around all of it.
What RegainFlow engineered
- Governed vocabularies anchored to published industry standards, so canonical meaning is defensible rather than invented in-house.
- A cross-reference engine mapping every raw value from every feed to its canonical meaning, which is the layer whose absence caused the fragmentation.
- A cascading resolution pipeline — deterministic rules first, then constrained clustering, then ML classification, then LLM adjudication — so the cheapest sufficient method decides each case.
- Firearms domain knowledge embedded directly into the clustering constraints and vocabulary design, which is what stops a plausible-looking merge from corrupting the catalog.
- Human judgment reserved for high-leverage decisions only, rather than spent on cases a rule already settles.
- A bronze, silver, and gold medallion framework, so raw capture, canonical mapping, and resolved truth stay separable and auditable.
- Independently labeled gold sets with controlled A/B experiments isolating normalization from resolution policy, so each layer's contribution is measured rather than asserted.
- Full lineage on every golden value, so any resolved attribute can be traced back to the sources and the decisions that produced it.
Selected artifacts
Reconstructed, not screenshotted.
The resolution cascade
- Deterministic rules
- Constrained clustering
- ML classification
- LLM adjudication
- Human judgment
What the client received
- A forensic audit locating the cause of catalog fragmentation
- Governed vocabularies anchored to industry standards
- A cross-reference engine from raw values to canonical meaning
- A four-stage cascading resolution pipeline
- Gold sets and controlled A/B experiments per layer
- Full lineage on every resolved attribute
Outcome
Fragmented votes collapse into consensus, attribute confidence rises, and catalog coverage expands to include source data that was previously unmapped. Every golden value carries its lineage, so a disputed attribute is a question someone can answer rather than argue about.
The work banks itself
Every resolution decision is kept, so the same judgment is never made twice. Human effort decays as the vocabularies fill in, while the catalog keeps improving as new feeds arrive.