Senior Recommender Systems Engineer who audits the feedback loop between recommendations and behavior before calling a model healthy. Kwesi covers the full recommendation-modeling surface: collaborative filtering versus content-based versus hybrid model selection, two-stage candidate generation and ranking architecture, cold-start handling for new users and new items, offline evaluation (NDCG, precision@k, recall@k) and its limits versus online A/B testing, and feedback loop or popularity bias diagnosis and mitigation. Who it's for ML and product teams running a recommendation system who need candidate generation/ranking architecture, cold-start strategy, and bias diagnosis tied to real catalog coverage, not an offline leaderboard number. Key capabilities Collaborative filtering vs content-based vs hybrid model choice based on interaction data density Two-stage candidate generation and ranking architecture sized separately for recall and precision Explicit new-user and new-item cold-start bridge strategies with a graduation point onto the standard model Offline NDCG/precision@k/recall@k results paired with what those metrics cannot confirm, plus an A/B design Exposure distribution and catalog coverage audits that catch popularity bias hiding behind strong accuracy How to use it Paste Kwesi's SKILL.md into your Claude Project Instructions (or any AI system prompt), then describe your Recommender Systems Engineer problem. Works with Claude, ChatGPT, and any AI chat. Under 2 minutes to install.
