About

I’m Jonathan Lim Siu Chi (@jonathanlimsc), sc are the romanized initials of my Chinese first name, 修齐.

I am passionate about AI and building scalable AI-driven products that impact real-world users.

On the more research side of things, I am broadly interested in improving the generalizability of ML models, looking at various AI sub-fields such as reinforcement learning, neuro-AI and multi-modal models.

In my 7 years of work experience, I have built:

• Built scalable, high throughput LLM and vector DB RAG systems that help thousands of customers generate pitch-perfect proposal answers from their uploaded documents and library entries
• Trained from scratch a GPT2-based LLM for novel molecule generation and co-authored Gotta be SAFE: A New Framework for Molecular Design
• Sentiment classification of customer service chat-logs, using RoBERTa trained from scratch on in-house datasets.
• Real-time Machine translation between >15 language pairs including English, Chinese, Thai, Vietnamese, Bahasa Indonesia, using FairSeq, deployed on GCP GPU instances, serving >200QPS per second, managed using Kubernetes.
• Recommendation system based on Graph Neural Network (LightGCN) to recommend brands to customers based on domain browsing behaviour extracted from telco logs.
• Convolutional Neural Network models for tumor detection and disease classification in MRI brain scans
• Real-time OCR pipeline based on Attention-OCR to extract text information from user identity cards, productionized with docker-compose. For monitoring model prediction distribution, I deployed the Elasticsearch-Logstash-Kibana (ELK) stack.