Johan Kok Zhi Kang

SPOT-PERP crypto asset trading bot. We did make a fair bit when FTX crashed lol.

I collaborated with an ex-JPMorgan senior trader to design a spot-perp arbitrage bot on multiple exchanges.

The bot looks at slippages in prices and executes a trade when the slippage is too large. In theory, this sounds easy but the implementation was a steep challenge.

I designed and optimized the trading logic so that most trades are profitable. I also built supporting infrastructure to automate the bot’s deployment for multiple spot-perp pairs.

Python library for building capacity scaled mixture of experts (MoE) models.

AdaptiveStream is a library that enables users to customize Rules defining when experts should be created and Policies specifying how these experts are created.

We enable fast scaling and inference over MoE models and are better than traditional continual learning frameworks focused on single model incremental learning.

Indexing machine learning models into a hierarchy for super fast knowledge access

Continuous learning models must balance between learning under concept drift and catastrophic forgetting.

We explore how we can index these trained models for fast knowledge retrieval and reboustness to catastrophic forgetting. We then developed a novel indexing scheme for each trained model and routing protocol for expert selection in logarithmic time complexity. More details can be found in our paper:

Johan Kok Zhi Kang, Sien Yi Tan, Bingsheng He and Zhen Zhang. Real Time Index and Search Across Large Quantities of GNN Experts for Low Latency Online Learning. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2023)

Graph partitioning algorithm refined using closed-loop feedback from model training performances

Most graph partitioning algorithms are deterministic. They partition a large graph based on topology.

We argue that it is more sensible to partition the graph with the objective of maximizing the training performances of machine learning models. We developed a novel algorithm that iteratively refines partition boundaries such that the final sub-graphs are optimized for model learning. More details can be found in our paper:

Johan Kok Zhi Kang, Suwei Yang, Suriya Venkatesan, Sien Yi Tan, Feng Cheng, and Bingsheng He. Dynamic Graph Segmentation for Deep Graph Neural Networks. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2022)

Deep learning, SQL language parsing for resource estimation

Prestroid helps developers figure out how much resources their SQL query needs. No need to worry about under provisioning resources, which causes queries to fail, or over provisioning, which leads to wastage in costs.

We analysed thousands of SQL queries and their resource consumption profile at Grab. We then improved on tree-convolutional models to develop state-of-the-art resource consumption forecast for a new SQL query. More details can be found in our paper:

Johan Kok Zhi Kang, Gaurav, Sien Yi Tan, Feng Cheng, Shixuan Sun, Bingsheng He. Efficient Deep Learning Pipelines for Accurate Cost Estimations Over Large Scale Query Workload. Proceedings of the 2021 International Conference on Management of Data (SIGMOD 2021)