Project case study
Product Recommendations and Bundle Builder
During my Home Depot internship, I turned purchase patterns into recommendation rules and a bundle-builder prototype. Merchants could inspect why an item belonged on a project list, and engineering received the data and rules behind those decisions.
My role: Data preparation, recommendation rules and product direction with AI-assisted implementation. A teammate contributed bundle rules; engineering built the retrieval prototype.
Watch the interactive walkthroughHow it works
Start with the job
The project need sets the context for a useful list.
Purchase pairs and professional-purchase frequency provide candidate relationships. Category filtering removes unrelated pairs before scoring.
Score related items
Related products receive a decision and a reason.
A local rule-based scorer assigns Recommend, Verify First or Do Not Recommend. Existing human labels are preserved.
Review recommendations
A merchant can inspect the list before using it.
The app exposes recommendation and bundle modes over the same underlying logic. Illustrative bundle values are not measured business results.
Hand off the rules
The decision contract survives the move to engineering.
Versioned inputs and outputs separate scoring from retrieval. The engineering retrieval layer can explain approved recommendations without replacing their governing rules.
Every recommendation should have a reason.
During my Home Depot internship, I filtered purchase-pair data, scored examples and developed rules for a project-list recommendation set. I then directed a local app that let users work with recommendations and bundles without writing code. The handoff separated the decision logic from any later AI explanation layer.
- Focus
- Local data processing, deterministic scoring, versioned schemas and retrieval contracts
- Status
- Recommendation data handed off; bundle-builder prototype
How it works
Local processing filters candidate product relationships, applies rule-based labels and retains reasons. The app exposes recommendations and bundles over that logic; versioned schemas define the engineering handoff and keep retrieval from silently changing the decisions.
A decision that mattered
The rules engine decides whether an item is recommended, needs verification or should be excluded. Retrieval and generative AI can explain the result but cannot silently override it. Human labels remain intact. Engineering used my data and rules to ground a retrieval prototype; I did not build that retrieval pipeline.
Status
The recommendation data and algorithm were handed off, and the app was shared for internal testing. Engineering’s retrieval prototype was a separate contribution; production deployment is not established here.