ML & Coding

With a solid foundation in multiple programming languages, my coding journey has been a continuous pursuit of knowledge and skills in computer science and machine learning.

I possess a deep understanding of the principles and applications of ML, having implemented traditional machine learning algorithms from concept to code entirely from scratch.

My passion in ML drove me further into the exploration of advanced concepts within deep learning. Here, I have not only simulated automatic differentiation but also explored more about understanding and applying JAX's tracing mechanism in learn-jax-with-code.

I developed tpux, a comprehensive tool to automate the process of configuring TPU VM and TPU Pod, based on user requirements. I also created the einshard, an einsum-like high-level API to shard and replicate arrays in JAX to simplify data and 1-D tensor parallel for machine learning models.

NLP & Languages

Languages are not only mediums of communication but also mirrors reflecting diverse cultures. My enthusiasm for language learning, having embraced English, Chinese, Japanese, and currently Estonian, naturally propelled me into Natural Language Processing. My journey in NLP, especially in Large Language Models, is driven by my curiosity about languages and how they can be processed and understood by computers.

My research work includes text classification tasks to finetune models like RoBERTa, T5, and FLAN-T5 with semi-supervised learning methods, contributing to an EMNLP workshop paper.

I was involved in a comprehensive survey on the trustworthiness of LLMs at ICML. My contributions to this survey included evaluating many aspects such as jailbreak, stereotype, disparagement, emotional awareness, etc. Specifically, I create benchmarks for LLMs with GPT API, where different types of data from the dataset served as prompts to generate results from various LLMs. I proposed evaluation metrics tailored to these datasets and assessed the LLMs' outputs using these metrics to measure their performance.

Through my open-source project mistral-jax, I implemented the Mistral 7B model from scratch using JAX with model architecture, tensor parallelism, and various generation methods, showcasing my technical expertise and engineering prowess.

I provided the tpu training example showing how to train a model on TPU for a simple machine translation task. This example used tpux and einshard, both tools developed by me, and served as an illustration of their effective utilization.

Gaming

My passion for gaming, with Divinity: Original Sin II standing out as a favorite. The passion for video games combined with my professional expertise in AI, motivated me to explore integrating AI into gaming. I have explored to introduced GPT as an agent to navigate text-based games. Besides, I also developed MahjongGenerator, an open-source project aimed at unifying and customizing Mahjong rules that can accommodate different variants.

While my journey into combining AI with gaming is just beginning, I am eager for collaboration and further exploration in this exciting field.

Others

Beyond my professional endeavors, I find joy and inspiration in a variety of hobbies that enrich my life and spark creativity. Walking serves as my best scenario for contemplation, while photography allows me to capture moments of solitude and the beauty of landscapes during my travels, rarely focusing on people but instead on the world around me. My musical journey has taken me from drums and erhu to guitar, and I'm currently learning the bass, excited about my future possibilities of blending AI with music to discover innovative sounds, beginning with bark-jax . Additionally, my open-source project schedule-main reflects my desire for simplicity in life event management, offering a straightforward way to track and analyze events and time in terminal without the burden of complex tools.

These interests not only provide a balance to my professional life but also inspire my work in technology, showing that innovation can stem from anywhere.