Roland Zhou

Machine learning engineer with 7+ years deploying end-to-end production ML systems across personalization, ranking, causal inference, and NLP. At Uber, modernized the Uber Eats home-feed ranker into a real-time Transformer-based listwise recommender, built Graph-RAG semantic shopping discovery over a product knowledge graph, developed the causal uplift layer for courier incentive optimization, and shipped the ML personalization behind Carbon Feed - a Driver-app surface serving 2.7M+ daily drivers and 30M+ daily card impressions. Previously at FactSet, built hybrid semantic search (BM25 + financial-domain BERT) and NER/entity-linking pipelines that structured millions of financial documents daily against canonical company identifiers.

Machine Learning Engineer - Uber AI / Delivery Marketplace / Eats Feed Intelligence at Uber (2020-12 – Present)

• Led the end-to-end modernization of the Uber Eats home-feed ranker from a batch-feature pointwise DCNv2 DeepCVR model to a real-time, target-aware Transformer with multi-task click/order heads and listwise slate scoring - owning dataset design, architecture, offline evaluation, serving, and staged rollout - lifting home-feed order conversion 1.4% and contributing ~$18M in annualized incremental gross bookings.
• Engineered target-aware Transformer sequence encoding that appends the candidate merchant to the user's chronological event sequence, letting multi-head self-attention re-weight behavioral history per candidate and capture in-session intent shifts aggregate features destroyed.
• Evolved the ranker from pointwise inference to listwise generative recommendation - scoring the entire merchant slate in a single Transformer forward pass - reducing per-candidate inference complexity to ~1/T and enabling slate-level reasoning over diversity, reorder balance, and merchant redundancy.
• Implemented position-bias correction with separated examination and relevance towers (presentation features train the bias tower; only relevance serves online), producing statistically significant gains in home-feed orders and orders per user.
• Optimized two-tower candidate retrieval with geographically constrained in-batch negatives and LogQ popularity correction, improving Recall@500 from 89% to 93% while shrinking the user-representation model ~20× via activity-based encoding.
• Built point-in-time-correct training pipelines on the Next Personalization Platform's event-sourced UserContext with shared online/offline Java FeatureExtractors, cutting behavioral-feature lag from 24+ hours to seconds while eliminating training-serving skew.
• Productionized PyTorch sequence models through offline ONNX-to-TensorRT builds and CPU/GPU-disaggregated serving with multi-hash embeddings and BF16 training, eliminating model cold-start conversion delays and materially improving per-node serving throughput.
• Architected a Graph-RAG shopping discovery system combining a heterogeneous product knowledge graph, GNN embeddings (GraphSAGE/R-GCN), HNSW + BM25 hybrid retrieval in Lucene Plus, schema-validated LLM query planning, and an XGBoost learning-to-rank layer - improving ranking relevance, recovering zero-result searches, and driving incremental grocery and retail gross bookings.
• Trained heterogeneous GNN embeddings (link-prediction, substitution, and contrastive shopping-mission objectives) with distributed Ray jobs on burst GPU capacity across AWS and GCP, powering candidate retrieval, product cold-start representation, and XGBoost ranking features.
• Designed grounded LLM query planning that converts natural-language shopping requests into JSON-schema-validated dietary, allergen, budget, and delivery constraints with deterministic-NLP fallback, sharply reducing unsupported generated attributes while enforcing hard allergy filters no score could override.
• Developed the ML personalization layer for Carbon Feed (named author on Uber's engineering blog), ranking earnings, promotion, and compliance cards via click-propensity models with near-real-time Kafka/AthenaX features in Michelangelo Feature Store, and automated its daily retrain-evaluate-deploy lifecycle through MLStudio with online scoring served via the Go-based Carbon Aggregator - scaling personalized ranking to millions of daily drivers and couriers.
• Built multi-treatment causal uplift models (CausalML, doubly robust estimation) predicting incremental courier supply and expected payout per courier–offer pair, feeding a Google OR-Tools CP-SAT optimizer that replaced first-in-first-out campaign contention with value-based allocation and near-full budget utilization.

Machine Learning Engineer - Research & Content Intelligence at FactSet (2019-01 – 2020-11)

• Built a hybrid semantic-search platform for financial research combining BM25, a financial-domain BERT bi-encoder over ANN passage indexes, entity-aware candidate fusion, and a BERT cross-encoder reranker - improving NDCG@10 by 13.4% and cutting median time-to-relevant-document 29% at sub-350ms p95 latency.
• Engineered hard-negative mining strategies (same company/wrong topic, same topic/wrong company, wrong fiscal period) and point-in-time training datasets from graded analyst judgments and click/dwell/export behavior with position-bias correction.
• Developed a domain-adapted BERT NER model (evolved from BiLSTM-CRF) extracting companies, securities, people, metrics, and transactions from filings, transcripts, and news, substantially raising exact-match F1 over the legacy rules-and-dictionary baseline.
• Built the entity-linking pipeline against FactSet Entity Master and Symbology - alias/ticker candidate generation, GBDT learning-to-rank with contextual BERT reranking, and publication-date-aware temporal resolution - delivering high-precision top-1 company linking and sharply reducing false document-company associations.
• Designed calibrated confidence thresholds, NIL detection, and an active-learning review workflow that sharply cut manual entity review while processing millions of document passages daily at a fraction of the prior enrichment latency.
• Implemented entitlement-aware retrieval and snippet generation so licensed broker research was filtered before any user-visible text, and shipped conversational research chatbot capabilities grounded in entity-linked content.

M.S. in Computer Science – Columbia University (2017 – 2018)