Keywords: Python, C++, Perception

* Improved the performance (PR curves and mIoUs) of the lidar semantic segmentation task, which is one of the main tasks of a multimodal perception model, by inputting noise point cloud in addition to non-noise point cloud and co-training the model with a noise-aware auxiliary task alongside the main tasks. The ML frameworks used were TensorFlow and JAX in Python.

Keywords: Python, Pre-training, Fine-tuning, Distributed Systems

* Rapidly acquired necessary expertise in LLM within 1 month despite working full-time in a different field (Self-Driving Car), and built an LLM distributed training pipeline with PyTorch, Microsoft DeepSpeed, and Hugging Face Transformers in Python, achieving successful training of a GPT-based 10B generative model on 160 Nvidia H100 GPUs via Google Cloud GPU cluster.

* Led a lecture on pre-training in an online LLM course with 4,000+ participants, accomplishing the highest Net Promoter Score (except guest lectures).

Keywords: C++, Python, Prediction, Motion Planning

* Decreased driver interventions in JP by 30% by debugging the trajectory planning modules in C++ with teams in the US, e.g. triaging log data, identifying root causes on large C++ codebases, articulating solutions, and managing vehicle experiments.

* Streamlined simulation scenario generation process for the motion planning team by 3x efficiency by creating automation scripts in Python that leverage Applied Intuition API and by leading cross-functional collaboration with teams in the US.

* Enhanced safety validation for Level 4 driverless autonomous vehicles by designing and implementing a prediction evaluator in C++ with software development life cycle best practices, e.g. object-oriented design, code standards, code reviews, and testing.

Keywords: C++, Python, Localization & Mapping, Perception

* Developed perception features in C++ by applying research theories to practical applications in a fast-paced culture, including a speed bump detector using lidar point cloud and probabilistic approaches (successfully delivered to a customer company) and a lane line visibility checker using lidar-camera sensor fusion and linear algebra (published in an international journal paper).

* Improved map data collection process for autonomous vehicles by 2x efficiency by creating a real-time mapping tool in C++.

* Mentored 4 lab students, resulting in 1 paper publication at IEEE IV and 1 patent filing, by providing hands-on guidance.