Availability
My earliest availability for work would be in late December of this year.
Currently, I quit job already so don't have any notice period.
Visa
I am planning to apply for a YMS visa application, which is mostly accepted because Sunak recently increased the quota from 100 to 5000 for South Korea. The visa status for next year would allow legal work.
Something meaningful
I created my own Knowledge Graph to share information, named Texonom. The Texonom dataset was used for my next QA AI project, LLaMa2GPTQ. This chatbot project can provide responses backed up by relevant documents retrieved from Texonom. I have optimized the LLaMa2 Model using GPTQ 4-bit quantization to ensure that everyone can use LLM in their local environment.
Felt compelled to build something
For individuals, productivity is strongly tied to the computer interface. I wanted to optimize it because I felt that the keyboard and mouse were bottlenecks. Standard input devices couldn't keep up with the speed of a normal person's thought process. When coding, Vim is a tool that enhances productivity, but its keybinding layout is not intuitive for everyone. I gamified the keyboard using arrow keys to control the keyboard and mouse very efficiently. This idea-driven development required an intense effort to learn new things.
AutoHotKey played a core role, and I also learned Electron, ultimately publishing it on Winget for everyone. This experience marked the first time I successfully productized software from end to end. It made me realize that I can create anything I want.
interpretability work
I use summarization AI when reading lengthy articles, finding it very useful. In the future, there is a high likelihood that filter bubbles will expand as people rely on summary AI instead of reading entire articles. If AI summarizes all information, potential vulnerabilities to AI hacking increase. If the summarization model is interpretable, it becomes easier to track why a particular summary was generated and detect instances of poisoning. RTSum is an AI designed for news article summarization, employing both extractive and abstractive techniques. RTSum breaks down sentences into relation triples and utilizes salient triples as intermediates to generate the final summary. The existence of intermediates provides interpretability and enhances responsiveness to potential attacks.
Want to work for interpretabillity
I view interpretability a critical aspect of both safety and alignment in artificial intelligence systems. An interpretable AI is not only transparent but also easier to steer, contributing to better alignment with intended goals. Achieving alignment is crucial, and interpretability provides the means to ensure that the AI system operates as intended.
In the pursuit of alignment, several methods for preference optimization exist, such as dataset manipulation, reinforcement learning, and prompt optimization. These techniques allow for fine-tuning and steering the AI toward desired outcomes, providing a level of control and alignment akin to human decision-making. An example of this is ControlNet, emphasizing that only controllable AI can achieve real usefulness.
Similar to humans, a lack of interpretability in black-boxed AI models can be risky. Interpretability, in this context, serves as a safeguard against unintended consequences and allows for a more informed and deliberate approach to AI development. Notably, the development of the sentiment neuron contributed to the advancement of GPT, and I believe that further interpretability research will lead to breakthrough models.
Even though the pursuit of alignment may not always result in a perfect model, having an explainable AI model offers the advantage of being able to choose the training path wisely. In summary, my experience and understanding of alignment and interpretability would be valuable for your team. I am eager to contribute by developing guardrail software for LLM.
I am particularly drawn to the prospect of working for {} due to its emphasis on user-centric product development. My experience with various vector databases and proficiency in handling large datasets align seamlessly with {}’s focus. The result-oriented development approach at {} resonates with my professional ethos. Having previously worked in AI product development with a similar technology stack, I am enthusiastic about the potential contributions I can bring to further enhance {}’s innovative projects.
Math
In the field of Math, symbolic regression is a mathematical technique that can be used to optimize real-world problems by approximating various sequences using mathematical equations. Process supervision paper such as LearnDojo and PRM800K can be used to delegate mathematical proofs to AI, which can help solve previously unsolved problems.
In the field of physics, simulation is important, but interpreting the data from the movement of numerous particles is also crucial. Combining particle motion with symbolic regression can be used to obtain physical laws or astronomical insights.
In the field of chemistry and biology, molecular property prediction like Drug Discovery is an important application. Machine learning could improve the low success rate of the discovery process.
Open-source tools like DeepChem and Cello should be more released for science development. By using these tools, scientists can improve their understanding of both fields and make new discoveries.
London
London, the city of my dreams, is a place where history meets modernity. Working here will be a dream come true, and I feel fortunate to be part of this vibrant city. London's rich history as the first modern city has always fascinated me, and I hope to experience its preserved breath. Moreover, it is a mentally stable place for me as my girlfriend will be studying at UCL from September. With its world-class education, finance, and cultural institutions, London is a hub of innovation and creativity. I am excited to explore all that this city has to offer and make the most of my time here.
SCIENTIFIC DOMAIN
As a software engineer, I have also developed several algorithms and ideas that have been published and implemented in various projects. One of my contributions is the multi-dimensional version of Chaikin's smooth algorithm, which I generalized and published as an NPM library. This algorithm is useful for smoothing out curves in multi-dimensional space.
In my previous job at Kakao Mobility, I invented a multi-ray real-time surface extraction algorithm in 3D space. This algorithm is useful for extracting surfaces from point clouds in real-time, which is important for applications for Map for autonomous driving. I also implemented an intensity-based lane extraction algorithm from 3D point cloud, which is useful for detecting lanes on the road and improving the accuracy of autonomous driving systems.
To optimize a local running Q&A AI, I quantized Large Language Model by GPTQ algorithm. This algorithm is useful for reducing the memory and computation requirements of large language models. Additionally, I came up with an idea of providing context documents that are similar to the input query using LangChain, which is useful for improving the inference time of language models. As a result of these contributions, I completed the personalized and locally executed LLM Chatbot product supporting CLI and Web UI. This chatbot is useful for answering questions and providing recommendations to users in a personalized and efficient manner.
Overall, my work has focused on developing algorithms and ideas that are useful for solving real-world problems in various domains, such as 3D mathematics and natural language processing.
Frontend
I am proficient in React and have extensive experience working on various projects. I skillfully utilize React Hooks where appropriate and possess a deep understanding of context and the rendering process. Additionally, I am familiar with frameworks like Next.js. Currently, I am managing a web application with over 10,000 pages using this framework, effectively employing static site generation (SSG) and server-side rendering (SSR) with fallback. My approach to coding in React is focused on creating highly readable code by adopting the atomic design pattern. I also emphasize writing optimized code, integrating state management only when necessary.
Furthermore, I have been using CSS for a long time and consistently keep up-to-date with the latest standards. I am proficient in creating layouts using Grid or Flexbox and have adopted recent techniques such as pseudo-classes and CSS Nesting. Fundamentally, I understand CSS rendering and strive to write code that is both minimal and optimized for performance. Of course, I also prefer Utility First CSS Framework like Tailwind CSS, and am familiar with seamless CSS integration using Shadcn.
Backend
I have worked on various Node.js backend projects, including implementing a web server accessed by 100 workers in real-time using Node.js. This work involved optimizing performance through the integration of a Redis cache. I am adept at enhancing backend servers with various functionalities, such as logging and caching, through middleware. Utilizing Node.js N-API, I have created an interface layer that allows the use of libraries from other codebases, employing C++ and Rust bindings. Regarding Python backends, I have a solid understanding of ASGI and WSGI concepts, which I am currently applying in a small-scale project.
Data visualization
While working at Kakao Mobility, one of my primary responsibilities was 2D map data visualization, particularly visualizing people's work activities on a map. I utilized InfluxDB, a time series database (TSDB), for logging the data and Grafana for its visualization. Specifically, I used the Grafana GeoMap Panel to manage workers' details and monitor their overall work in real-time.
Another significant task involved displaying large volumes of 3D data. I gained state-of-the-art (SOTA) experience in displaying and interacting with tens of millions of PointCloud data on the web in real time using Potree, where I developed various skills. Additionally, I visualized tens of thousands of 3D vectors in geojson format.
natural language processing, sentiment analysis, and machine learning
My career is primarily focused on NLP (Natural Language Processing) and machine learning. I worked on 'RTSum', an academic paper dedicated to AI summarization. In this project, text preprocessing involved sequentially segmenting it into sentences and semantic triples, followed by training an abstractive summarizer to convert this information into summaries. Currently, I am working on a paper titled 'ReSRer', which aims to optimize the method of providing prompts within the context given to question-answering AI. During this project, 6 million Wikipedia documents were segmented into 256-token units, resulting in 21 million Wikipedia passages that were then indexed into a vector database.
I am adept at asynchronous processing and utilizing multiple GPT instances while managing large datasets. A notable example of this is optimizing the Sentence Transformer to process 21 million data points within a single day, leveraging Text embedding inference. For Large Language Models (LLMs), I have successfully reduced memory usage by four times through the application of GPTQ.
structured data and unstructured data analytics
In my experience with unstructured data, I covered the entire English data on Wikipedia. Throughout this project, documents were segmented into passages of 256 tokens each, with anomalies identified based on text length. I developed a statistical approach by categorizing both token and text lengths in exponential ranges, such as from 1024 to 2048 tokens. Anomalies were detected in cases where documents contained excessively lengthy JSON structures or repetitive HTML sequences lacking whitespace. After removing these anomalies, I successfully constructed a dataset specifically designed for AI-based question answering using the RAG technique. This dataset is accessible here: https://huggingface.co/datasets/seonglae/wikipedia-256
Additionally, I have specialized in structured data analysis, with a particular focus on geography. While at Kakao Mobility, I was tasked with analyzing extensive line and surface vector data in geojson format, processing batches of approximately 100,000. This detailed analysis was essential to identify instances where property relationships were connected, but the actual geographical coordinates were misaligned, or the reverse. To tackle this challenge, I employed a sophisticated matching algorithm that methodically progressed from smaller to larger data classes. This strategic methodology enabled us to detect anomalies accurately within the designated delivery period, thereby ensuring the provision of precise and reliable data.
data analytics and statistical analysis
I was tasked with analyzing the usage statistics of a web application utilized by 100 employees. To achieve this, I developed a feature to track and log task creation and modification details for each employee. These records were meticulously catalogued as a time series in a Time Series Database (TSDB). Administrators, when accessing the statistics interface, used the InfluxDB Flux language to query data. This functionality allowed them to examine the volume of modifications and creation tasks completed by each individual within a specified timeframe. This process greatly assisted managers in devising more effective worker management strategies and implementing competitive ranking rewards.
In addition, I played a key role in error log collection from both the frontend and backend of the application. By utilizing Grafana Alerting, I established a notification system that alerted administrators via Slack when the error count surpassed a pre-set limit within a certain timeframe. Moreover, my efforts extended beyond GitOps as I successfully integrated ChatOps into our operational structure. This integration proved vital in fostering a unified intra-team workflow, significantly boosting our team's overall productivity.