Developed a custom callable proposer backend for speculative decoding in the jeejeelee/vllm repository, enabling users to define their own proposer classes for token generation workflows. This feature focused on enhancing extensibility and experimentation within speculative decoding pipelines, allowing for greater flexibility and potential performance tuning. The implementation emphasized modular backend design and plugin-like extensibility, leveraging Python and machine learning principles to support advanced AI development needs. Throughout the process, code quality and maintainability were prioritized through signed-off commits and clear pull request communication, with a strong emphasis on collaborative development practices and comprehensive testing to ensure robust integration.