Projects

Enterprise GenAI platform serving 15,000+ CBA employees. Enables teams to build and consume LLM-powered applications for content generation, knowledge retrieval (RAG), and process automation.

• › Defined API interfaces and system boundaries through stakeholder engagement
• › Introduced PromptFlow as the Low-Code GenAI orchestration framework
• › Delivered RAG app backed by Azure AI Search + GPT-4, deployed on AWS
• › Extracted reusable patterns into internal Python libraries
• › Defined team Python coding and deployment standards

Designed and operated a production-grade on-premises data platform from scratch — replacing cloud spend with a Kubernetes-native stack that the whole analytics org runs on.

• › Stood up Spark 2.x → 3.x cluster on Kubernetes with Ceph as S3-compatible storage
• › Set up JupyterHub + SparkMagic + Livy for self-service PySpark by Data Analysts
• › Deployed and managed Airflow for batch ETL scheduling and monitoring
• › Integrated Kafka → Spark Streaming for near-real-time analytics
• › Evaluated Delta Lake vs Parquet for raw data storage

Collaborative Filtering recommender built on Spark MLlib to surface relevant software features to users — driven by behavioural telemetry analysis.

• › Prototyped using matrix factorisation (ALS) via Spark MLlib
• › Ran random search + heuristic global hyper-parameter optimisation
• › Full feature-engineering pipeline from raw user activity logs

Production chatbot for enterprise customers — NLU pipeline combining Rasa with SpaCy, BERT, and TensorFlow components.

• › Led architecture and development of the data ingestion pipeline enabling SMB market entry
• › Prototyped BERT semantic search for sales FAQs
• › Introduced and presented Deep Learning (RNNs) company-wide
• › Built internal DevOps Hubot for deployment automation

An RNN-based natural language generation model that writes classical Tang-dynasty poetry. Deployed as a WeChat chatbot with zero marketing — tens of thousands of users, hundreds of thousands of poems generated.

• › Implemented in Python / TensorFlow with LSTM cells (open-sourced on GitHub)
• › Compared GRU vs LSTM cells via A/B testing
• › Full solo operation: product, marketing, analytics, dev, devops

Item-based Collaborative Filtering recommendation engine processing billions of records per run on Amazon EMR. A/B tested 36% better CTR than the third-party incumbent.

• › Built with Java / Apache Mahout on on-demand EMR Hadoop clusters (EC2 + S3)
• › Applied business rules: metadata boosting + soft recency filtering
• › Served the full SMH and The Age newsletter audiences