← Selected work

LG Electronics Production ML

Who to reach. What to recommend.

A connected suite of customer and product models that turns purchase signals into audience priorities, relevant recommendations, and informed outreach.

Predictive MLRecommendation systemsModel evaluationAI-assisted ranking
My contribution

Technical leadership and hands-on delivery, from model design and evaluation to integration with business workflows.

Delivery status

Customer models in production

Four questions. Connected answers.

The customer models shape the audience and offer. Product alternatives address what to suggest when availability changes.

The business problem

A useful customer decision needs more than a score. Teams need to know which customer to prioritize, what product is relevant, what price level fits, and what to offer when that product is unavailable.

What I built

Led the development of a connected customer intelligence product at LG, bringing next-purchase recommendations, customer value, and category spending into a shared decision workflow.

Turned model outputs into usable signals for retention and outreach teams: ranked categories, purchase readiness, expected value, and category-specific spending preferences.

Extended the work to product alternatives, combining catalog context, availability rules, and AI-assisted ranking to prepare explainable recommendations for downstream channels.

How I evaluated it

Evaluated customer predictions against later purchases using information available at prediction time. Compared model combinations with simpler audience-selection baselines, and assessed product alternatives through automated checks and independent AI review.

Built to operate

Built repeatable scoring and validation into the product. Publication checks help keep incomplete or invalid results from reaching downstream workflows; retraining gates are designed to compare replacements before promotion.

What the work enables

Delivered customer models in production and an operational product-alternatives pipeline. Historical evaluation supports better audience prioritization than broad targeting baselines; customer-facing activation and incremental sales require separate validation.

Selected details are generalized to respect confidentiality. Internal metrics, customer records, and proprietary implementation details are omitted.

Let’s talk about the work