35% Recommendation & Personalization Modeling
Own the design and development of the core recommendation models that turn user-saved product data into a personalized feed. Develop multi-signal models spanning brand affinity, category, color/visual attributes, fit and sizing, price sensitivity, and trend. Select and justify approaches across collaborative filtering, matrix factorization, content-based, and hybrid/neural methods (e.g., two-tower and other embedding models), and know when each applies. Build product and user embeddings that capture semantic similarity across the catalog and power candidate generation and retrieval. Design cold-start strategies that produce high-quality recommendations for new users and newly ingested products with little or no behavioral history.
25% Experimentation & Measurement
Define what “good” personalization means and how it is measured. Establish rigorous offline evaluation (ranking and relevance metrics, sound holdout design) and connect it to online outcomes. Design, run, and read out A/B and multivariate experiments, and translate results into clear product and business decisions. Bring statistical discipline — sound experiment design, awareness of bias and confounding, and honest interpretation — so the team can trust which changes actually move engagement.
20% Data, Signals & Feature Understanding
Develop deep intuition for Picksy’s product catalog and user signals. Turn implicit behavior (saves, clicks, dwell, shares) and catalog attributes into meaningful model features, writing SQL against BigQuery to pull, join, and shape raw data into training/evaluation datasets. Apply NLP and computer-vision techniques — including modern embedding and LLM-based approaches — to extract structured attributes (category, color, material, fit) from unstructured product descriptions and imagery, and to enrich sparse catalog data. Partner with data engineering on data quality, freshness, and coverage as the catalog scales from hundreds of thousands toward tens of millions of products.
20% Full-Cycle Ownership & Productionization
Take models from prototype to production yourself. Write clean, production-quality code and deploy into the existing MLOps pipeline (feature store, training, serving, monitoring) rather than building infrastructure from scratch. Own model performance in production: instrument it, watch for drift and degradation, and iterate as behavioral signals accumulate. Document models, features, and decisions clearly, and collaborate closely with the MLOps, engineering, and product teams to integrate the model layer into the live product.