System Diagnostics: Root cause the bottlenecks of the current search and recommendation systems and quantify the defects across the entire search funnel (query rewriting\retrieval\ranking).
State-of-the-Art Solutions: Propose and architect end-to-end solutions that leverage advanced ML, DL, and Generative AI/LLMs to upgrade traditional search components into AI-native architectures.
Core Funnel Optimization with LLMs:
Query Understanding: Build next-generation query understanding services (intent detection, query rewrite, semantic expansion, entity extraction) utilizing fine-tuned LLMs.
Candidate Retrieval: Design and scale semantic/dense retrieval systems, combining traditional vector search (ANN) with generative retrieval techniques.
Ranking & Reranking: Develop and optimize LLM-based listwise/pairwise reranking models, and tackle performance trade-offs (latency, cost, throughput) for real-time serving.
Production & Evaluation: Implement solutions, optimize offline/online inference performance (using model distillation, quantization, vLLM/TensorRT), and conduct rigorous A/B tests.
Technical Leadership: Be a thought leader, mentor senior engineers, and gain consensus across cross-functional organizations (Infrastructure, Platform, Product) to drive the search AI roadmap.
LLM-Powered Query Understanding: We train and fine-tune in-house LLMs to perform multi-task query tagging, real-time query rewriting/expansion, and precise intent categorization.
Generative & Semantic Retrieval: We use deep representation learning, two-tower models, and LLM-generated semantic embeddings combined with vector databases (e.g., Milvus, Vespa) for hyper-accurate candidate retrieval.
LLM Reranking & Distillation: We explore listwise and pairwise LLM reranking models to evaluate query-item relevance, while distilling their reasoning capabilities into smaller, sub-millisecond models for real-time serving.
Deep Conversion & Personalization: We build Deep Learning-based conversion prediction models ($CVR$) incorporating real-time user state, query intent, and multi-modal product representations.
Graph-based Relationships: We leverage graph neural networks (GNNs) and knowledge graphs to ground our LLMs, exploring latent relationships between query sessions and catalog documents.