Tweak is 'AI Photo Community'. By inputting training data of 8 to 20 images, the service learns to generate photos using the user's face.https://apps.apple.com/vn/app/tweak-ai-photo-community/id1664306622I performed two major tasks:- infrastructure/architecture development for personalized stable diffusion pipeline: I designed an architecture for inference using cloud GPUs. It was important to stabilize the latency distribution while keeping costs low. To reduce costs, I used Spot Instances, which are available at approximately one-third the price of other instances but can be randomly reclaimed by providers. Developing safety measures was essential to using Spot Instances effectively.- stable diffusion training/inference pipelines development: SD 1.5, SDXL 1.0, Text2Image, Image2Image, Inpainting pipelineI also worked on other infrastructure-related tasks.- I built ArgoCD in a Kubernetes environment to adhere to the Single Source of Truth principle. This will help to manage and synchronize application deployments across different environments, ensuring consistency and reducing configuration drift. ArgoCD is a powerful tool for continuous deployment and maintaining a reliable infrastructure.- To achieve stable deployments, I implemented the Canary deployment strategy using Argo Rollouts. This approach allows for gradual rollout of new versions of the application to a subset of users or instances. By monitoring the performance and collecting feedback, we can make informed decisions and ensure stability before fully deploying the new version. Argo Rollouts provides a reliable and automated solution for managing Canary deployments in a Kubernetes environment.

AnnoWiz is a tool that helps to generate high-quality training data more efficiently based on semi-automatic data labeling technology. It automates the bounding box and polygon segmentation tasks with 70% to 99% accuracy depending on the class, reducing the time needed for AI training data processing by up to 90%. This is particularly useful for industries that require a lot of data preprocessing.At Annotation-AI, I worked on the following projects:- Worked on Triton Serving, an open-source software stack for AI inference workloads. This allowed for efficient deployment of deep learning models in production.- Implemented TensorRT quantization to optimize throughput and memory usage of deep learning models.- Developed Segmentation Similarthings, which applies machine learning to find similar objects while adjusting the number of inferences. This is a more efficient way to segment similar objects in an image compared to segmenting all objects in an image which requires 1,024 inferences per a single image.- Wrote a technical blog on gRPC load balancing, which explained how to implement load balancing in gRPC using Kubernetes and Traefik.- Optimized Python FastAPI to improve its performance and efficiency.- Conducted load testing and profiling to identify and fix performance issues in the system.- Developed a Golang Echo API for more light and efficient API.- Developed a logging system architecture in Kubernetes with Loki, Promtail, and Grafana. This allowed for efficient monitoring and troubleshooting of the system.

Spacewalk analyzes the characteristics of the environment that developers want to develop by combining various data such as land, transportation, surrounding prices, and building regulations. It derives development scenarios optimized for the given environment using design strategies acquired by artificial intelligence through deep reinforcement learning.At SpaceWalk, I worked on the following projects:- Accelerated the reinforcement batch inference speed through unlimited parallel processing using AWS and Kubernetes. This involved building a scalable and efficient infrastructure for reinforcement learning models.- Built a workflow pipeline for AI services using Argo Workflows. This allowed for efficient and automated execution of machine learning workflows.- Built ArgoCD for CI/CD, which is a continuous delivery tool that automates the deployment of applications to Kubernetes clusters.- Built ArgoCD Vault for protecting secrets, which is a secure and scalable way to manage secrets in Kubernetes.Conducted a seminar on readable code and code review for machine learning engineers. This helped to improve the quality of code and collaboration among team members.- Used Kubernetes in AWS to build a scalable and reliable infrastructure for machine learning workloads.

Markinarocks is an AI startup that specializes in providing AI solutions and enterprise ML Ops platforms (Markinarocks Runway) for industries such as semiconductors, automobiles, manufacturing, and energy. It was selected as one of the "2023 Top 100 AI Companies" by CB Insights, a global market research firm. It is the only Korean company among the companies selected this year.At Makinarocks, I worked on the following projects:- Developed an anomaly detection service for robotics based on autoencoder. This involved training deep learning models to detect anomalies in sensor data from robots. [News]- Developed a residual autoencoder and resolved the noisy gradient problem with warm-up and proved it with mathematical expressions. This improved the performance and efficiency of deep learning models.[Blog Post]- Developed a regression test pipeline for machine learning models. The pipeline automatically runs train and evaluate whenever a commit is pushed with GitHub self-hosted runner. The runner is running on the company's own GPU, which allows for efficient and scalable testing of machine learning models.

VUNO is a medical AI company that provides AI-based fundus image reading support solutions for medical professionals. Its core AI technology "Method for supporting fundus image reading for a subject and device using the same" has been patented in the United States. VUNO's technology considers the anatomical characteristics of the fundus to provide user convenience when reading fundus images using AI. It automatically finds the center of the macula and optic disc in fundus images and outputs diagnostic information in each segmented area based on those centers. This provides information on the fundus segmentations and allows for intuitive confirmation of the location-specific features of major findings.At VUNO, I worked on the following projects:Developed a reliable classifier using Class Activation Map. This allowed me to explain how the model predicts the data, which is important for transparency and interpretability of the model.Dockerized production-ready ML models. This involved packaging the models and their dependencies into Docker containers, which made them portable and easy to deploy across different environments.

- Review HCI paper and conduct experiment.- Experiment Subject: How Individual and virtual agent affect user's sport enviorment.