ML Researcher, NinoX¶
Summary
- Developed the Online Bayesian Inference (OBI) algorithm for the Collaborative Topic Regression (CTR) model—a hybrid system combining Collaborative Filtering and Topic Modeling—using Python with robust OOP design, while leveraging NumPy and SciPy for scalable, real-time recommendations.
- Achieved a 23% boost in computational efficiency by speeding up the MCMC sampling core with Cython.
- Established CI/CD pipelines with TravisCI for automated unit tests and a docs-as-code workflow using pytest and Sphinx, streamlining both development and documentation processes and ensuring consistent release quality
NinoX initially focused on IoT sensors and logistics management. The team was exploring new business opportunities when they discovered my proof-of-concept project. This led to the idea of building and productizing a recommendation system.
At NinoX, I developed the Online Bayesian Inference (OBI) algorithm, a hybrid system combining Collaborative Filtering and Topic Modeling. Using Python with a solid object-oriented design, I leveraged libraries like NumPy and SciPy to create a scalable, real-time recommendation engine. Additionally, I optimized the MCMC sampling core with Cython, achieving a 23% improvement in computational efficiency. To ensure smooth development and consistent quality, I established CI/CD pipelines with TravisCI, incorporating automated unit testing with pytest and a docs-as-code workflow using Sphinx.
This role marked my first experience with object-oriented programming (OOP) design patterns and using Cython to accelerate Python code. I also gained insights into variational inference, a concept I hadn’t encountered during my graduate studies. Beyond technical skills, I learned from my manager about algorithm development, CI/CD pipeline design, and the importance of efficient workflows. Most importantly, this experience gave me the confidence to tackle complex challenges and deliver impactful solutions.
See here for the original project documentation.
CI/CD Pipeline for OBI Algorithm Development