Machine learning engineer · Recommendation systems · Singapore
Four years building recommendation and personalisation systems at ByteDance Singapore — from offline feature engineering on Spark to real-time serving for 14M daily active users.
Case study
The two-tower retrieval model behind the recommendations surface, from question to production.
Could a two-tower retrieval model replace heuristic candidate generation and improve downstream ranking quality?
Designed and trained a FAISS-backed two-tower model on 90 days of interaction history, then shadow-tested it against the heuristic for three weeks.
Adopted in production: recall@100 up 18%, contributing to a 22% CTR lift on the recommendations surface.
Systems & models
Six pieces of work across three roles at ByteDance Singapore.
Profiled inference jobs, found 40% idle GPU time and moved batching to Triton dynamic batching with auto-scaling on Kubernetes.
Serving cost down 29%; p99 latency from 72 ms to 38 ms
Lead engineer for moving the feature store from Redis Cluster v5 to v7.
Zero-downtime cutover for 14M DAU traffic
Rewrote the Spark feature pipeline to use delta tables with incremental processing, cutting full-scan jobs.
Training cycle from 11 hours to 3.5 hours — same-day model freshness
Content-based embeddings blended with collaborative signals through a gating network, for users with sparse history.
+9 pts new-user 7-day retention in an A/B test across 620,000 users
Built and maintained MLflow experiment tracking infrastructure.
Model cards standardised for 12 production models
Watch-time regression with position-bias correction on logged data, evaluated offline with NDCG@10 and selected for an A/B test.
Average session length up 7% over a 4-week test
Technology stack
An ML engineer who works across the full model lifecycle — offline features and training on Spark through to low-latency serving with Redis and Kubernetes.
Luke joined ByteDance in Singapore as a graduate data scientist in 2021, became a Machine Learning Engineer in 2022 and a Senior Machine Learning Engineer in 2024. His models move engagement metrics for tens of millions of daily active users.
Beyond shipping models, he wrote an internal wiki on feature engineering best practices that three squads adopted. His wider skills include recommender systems design, A/B testing and experimentation, distributed training (DDP) and feature store architecture.
Experience
ByteDance, Singapore
Two-tower retrieval model (+18% recall@100), Triton serving (29% lower cost, 38 ms p99), Redis feature store migration.
ByteDance, Singapore
Incremental Spark feature pipeline for daily retraining, cold-start ranking branch, MLflow tracking and model cards.
ByteDance, Singapore
Watch-time prediction with position-bias correction; an internal feature engineering wiki adopted by 3 squads.
Education
Hong Kong University of Science and Technology (HKUST)
First Class Honours. Dean's List 2019–2021.
Contact
luke.zhang@email.com
Fictional candidate — contact details shown as plain text only.