shawnwang

Senior Backend & Smart Contract Engineer

Experience: 6 years

Yearly salary: $12,000

Hourly rate: $60

Nationality: πŸ‡ΈπŸ‡¬ Singapore

Residency: πŸ‡ΈπŸ‡¬ Singapore


Experience

Senior Engineer
OKX
2024 - 2025
Paragon, a spot matching engine on Aeron Cluster (Raft): three nodes replicate the command log, sharded by trading pair; events are persisted and fanned out to counters, market data and push. Responsible for feature development and production stability. - Matching engine development: features and production issues across command sequencing and dedup, ART-based order book matching, event storage and fan-out, snapshots and recovery. - Correctness: matching is a deterministic state machine, so historical commands are replayed to diff output between engine versions, and events are continuously compared across the three nodes.
Senior Engineer, Search & Recommendation
Shopee
2020 - 2024
ImageSearch - Content-Based Image Retrieval (PIC) E-commerce image search serving 7 Southeast Asian markets, ~120M products and 10K+ QPS; owned backend architecture, development and offline indexing. - Online pipeline: five-stage pipeline (embedding, multi-channel retrieval, dedup and cache, ranking, blending), each stage its own service; new sources or models plug in without touching the main path. End-to-end p99 < 800ms. - Multi-channel retrieval: i2i / u2i / ads in parallel, blended by slotting and top-k; online HNSW over all products with precomputed neighbors as a low-latency fallback for hot items. - Vector search engine: Elasticsearch 8 + HNSW over 120M+ 128-dim vectors; image MD5 as document key; per-country indices and category routing to a single shard. Vector retrieval p99 < 100ms. - Indexing: Hive + Flink + Kafka full + incremental indexing, idempotent by document key and replayable; shards sized and scaled per country. - Feature consistency: detection, feature and multimodal models share one implementation offline and online. - Multimodal: helped the algorithm team evolve retrieval from vision-only to image-text (image-to-title, user query images, cross-attention fusion); owned training data, A/B tests and rollout. EngineProxy - Retrieval & Ranking Engine Rebuild (Main Developer) - Moved query processing, retrieval, ranking and ads execution from the business layer into the engine, decoupling product and algorithm iteration. - Config-driven chain-of-responsibility pipeline for preprocessing, parallel retrieval, merging and ranking; compared Vespa and Elasticsearch to inform engine selection.

Skills

golang
english
chinese-mandarin