[{"slug":"llm-agents","order":1,"title":"LLM Agents","nav_title":"LLM Agents","short":"Skill retrieval, semantic-ID diagnostics, NL2SQL evaluation, and LLM serving systems.","headline":"Agent capabilities as reusable system assets","summary":"I study how LLM-based systems expose, retrieve, evaluate, and govern capabilities when they become reusable infrastructure.","question":"How can agent capabilities be selected, inspected, and evaluated before they become production dependencies?","scope_headline":"Evaluation before dependency","scope_text":"Skill retrieval, semantic identifiers, NL2SQL benchmarks, and LLM serving systems are treated as capability surfaces that should be inspected before they are trusted.","stats":[{"value":"4","label":"related publications"},{"value":"2026","label":"recent work on skill retrieval and semantic-ID diagnostics"}],"cards":[{"title":"Skill retrieval","text":"Evaluating ambiguity when multiple skills expose similar capabilities and must be selected reliably."},{"title":"Semantic-ID diagnostics","text":"Inspecting semantic identifier mappings before downstream recommendation training."},{"title":"NL2SQL and serving","text":"Benchmarking business intelligence services and scaling LLM inference systems."}]},{"slug":"recommender-systems","order":2,"title":"Recommender Systems","nav_title":"Recommender Systems","short":"User reactivation, generative recommendation, dynamic retrieval, behavior modeling, and trustworthy prediction.","headline":"Recommendation under changing behavior and scale","summary":"I work on recommendation problems where user behavior, retrieval spaces, and deployment constraints change over time.","question":"How can recommendation systems model returning users, retrieve efficiently, and remain reliable under industrial-scale constraints?","scope_headline":"Signals, retrieval, and prediction","scope_text":"User reactivation, generative recommendation, dynamic retrieval, graph embedding, counterfactual watch-time prediction, and efficient CTR modeling form the main technical line.","stats":[{"value":"6","label":"related publications"},{"value":"2026","label":"recent work on user reactivation, generative recommendation, and dynamic retrieval"}],"cards":[{"title":"User reactivation","text":"Recalibrating recommendation when returning users have pre-gap history but no observed behavior during long inactivity gaps."},{"title":"Generative recommendation","text":"Representing long histories, structured intent, and preference-aware generation."},{"title":"Dynamic retrieval","text":"Efficient item retrieval and graph embedding under changing user and item behavior."},{"title":"Trustworthy prediction and efficiency","text":"Counterfactual watch-time prediction and efficient CTR modeling for deployment."}]},{"slug":"data-mining","order":3,"title":"Data Mining","nav_title":"Data Mining","short":"Weak supervision, robust learning, text mining, analytics, and field experiments.","headline":"Reliable learning from limited and noisy evidence","summary":"I use data mining methods to learn from weak labels, sparse evidence, operational logs, and field experiment settings.","question":"How can learning systems extract reliable signals when labels, behavior, and operational data are incomplete or noisy?","scope_headline":"Learning from imperfect evidence","scope_text":"Weak supervision, robust learning, topic modeling, continual graph learning, and digital business experiments form the modern method line.","stats":[{"value":"13","label":"related publications"},{"value":"2","label":"digital business field-experiment papers"},{"value":"2010s","label":"early sequence-analysis work"}],"cards":[{"title":"Weak and robust learning","text":"Methods for learning when labels are scarce, noisy, or created through weak supervision."},{"title":"Text and graph mining","text":"Topic modeling, continual graph convolution, keyword graph learning, and data augmentation."},{"title":"Digital business experiments","text":"Live-streaming analytics and AI assistance studied through field experiment settings."}]}]