Original practice scenario inspired by Reliance Industries Limited's public business context; this is not a reported interview question or a description of its internal systems. A fictional diversified group wants to match product records from several retail catalogs. Identical names can describe different pack sizes, while one product may have several names. Design an entity-matching evaluation.
Read model answer +
Define a match as the same sellable product variant, including attributes such as pack size and unit. Text similarity alone cannot resolve identity when a misleading near-match changes what the customer receives.
Work through it
Build labeled pairs with exact matches, spelling variants, and difficult negatives differing in size or formulation. Separate product families across training and evaluation splits.
Use identifiers and normalized attributes where reliable, then score candidate pairs. Evaluate candidate retrieval separately from the final match decision.
Report precision at the automatic-merge threshold and recall including missed candidates. Route ambiguous cases to review and record reversible merge provenance.
Tradeoffs to discuss
Higher automatic precision usually reduces coverage. A review queue can recover coverage, but its workload and correction rate belong in the evaluation.
Common pitfalls
Random pairs make negatives unrealistically easy.
Merging two records is a business action; a high embedding similarity is not sufficient authorization.
Original practice scenario inspired by Reliance Industries Limited's public business context; this is not a reported interview question or a description of its internal systems. A synthetic retailer stores baskets, line items, and promotion applications. Several promotions can touch the same basket. Explain how to compare basket revenue for promoted versus unpromoted baskets while preserving one contribution per basket.
Read model answer +
Classify and aggregate at the basket grain before summarizing the promotion groups. Joining line items directly to promotion applications multiplies revenue whenever a basket has more than one application.
Work through it
Sum signed line-item amounts per basket, including the defined treatment of returns and discounts. Independently derive a boolean indicating any qualifying promotion application.
Join those one-row-per-basket results to completed baskets. Group by the promotion flag and report basket count, total revenue, and average basket revenue.
Test a basket with two items and two promotions, an empty basket, and a refunded basket. Reconcile total revenue to the original basket aggregate.
Tradeoffs to discuss
The result describes associated basket values, not incremental promotion effect. Customer selection and campaign targeting can confound the comparison.
Common pitfalls
SUM(DISTINCT amount) can remove legitimate equal-priced items.
Counting application rows as baskets distorts both the denominator and group assignment.
