Pipeline Architecture
3-stage ML pipeline: Retrieval → Ranking → Re-Ranking
Interactive Demo — Trace a Request Through the Pipeline
Two-Tower Model
Parameters6.5M
Embedding dim256-d
LossSampled softmax
TrainingGPU + AMP
Best epoch5 / 17
LightGBM LTR
ObjectiveLambdaRank
Features32
Val NDCG@100.4169
Train users5,000
Candidates/user100
MMR Reranker
Lambda0.7
Freshness weight0.05
Max same category3
SimilarityCosine
FallbackCategory-diverse