kuairand

End-to-end multi-task recommendation system with SASRec, MMoE, real KuaiRand evaluation, FastAPI serving, and an interactive model lab.

Causal analysis of random exploration in short-video recommendation using KuaiRand-27K

Evidence-first end-to-end recommender system on KuaiRand: retrieval, LR warm-start residual ranking, constrained reranking, FastAPI and interactive demo

Generative sequential recommendation: Meta HSTU on KuaiRand short-video logs, with a dual-stream eval protocol and a 0.56M–7.3M capacity sweep.

End-to-end short-video growth & recommendation analytics: PostgreSQL warehouse, LightGBM model, Streamlit dashboard — 2.6M KuaiRand-Pure interactions

Autonomous ML research agent that improves ML pipelines through typed graph mutations, ablation-guided search, leakage checks, and statistical validation.

Unbiased offline evaluation on real Kuaishou logs: ranking survives randomized exposure, calibration does not, and no single calibrator fixes both.