Worked on the jeejeelee/vllm repository to deliver a feature enhancing ColBERT’s backbone flexibility, enabling support for a range of non-standard BERT architectures. This development focused on expanding compatibility and usability, allowing broader deployment scenarios and easier integration with diverse model backbones. The implementation was carried out in Python, leveraging skills in API development, machine learning, and model development, with an emphasis on natural language processing techniques. The code was committed with attention to quality and maintainability, laying the foundation for future cross-architecture support and facilitating adoption by projects requiring adaptable language model infrastructure without introducing new bugs.