About Me
Machine Learning Engineer with 4+ years of work experience.
If you'd like to know more about me, click then type `import jinsung`
Interest
Machine Learning Engineering (ML Serving and MLOps)
Techical Skills
Languages: Python, Typescript
Backend Frameworks: FastAPI, NestJS
Deep Learning Framework: PyTorch
Work Experience
Dable (Subsidiary of Yanolja) | South Korea (Jul 2022 - Present)
Machine Learning Engineer (전문연구요원 전직)
AI Team (online ads system)
Server Side: (focused on optimizing the p99 latency and throughput)
- Maintained onnx-based ML model inference server processing 100k+ requests per minute.
- Implemented redis-cluster-based online feature store for mobile app DSP.
- Improved p99 latency and throughput of inference server, based on cpu profiled data.
- Reduced memory usage by 5X using shared memory, resulting in cost savings.
- Decreased model preparation time by 5X through implementing a model caching storage.
- Enhanced ML server observability with custom Prometheus metrics and Grafana dashboards.
- Implemented centralized logging system using CloudWatch Logs and Athena.
- Automated server code deployment, loadtesting, linting, code review.
- Refactored inference server structure for improved efficiency and manageability.
- Restructured inference architecture to support model ensemble methods.
Infra Side: (focused on optimizing the pipelining cost)
- Led migration of server applications to EKS cluster.
- Maintained Airflow DAGs for model pipeline (sensors, training, calibration etc).
- Improved pipeline job observability and fixed failing Airflow tasks.
Model Side: (focused on optimizing RoAS)
- Analyzed data of underperforming models to optimize budget allocation.
- Discovered and implemented new features, improving offline AUC of ML models.
- Developed new calibration model logic, increasing RoAS by 3X.
- Conducted A/B testing on calibrated pCTR model.
Vision Team (offline ads system)
- Involved in both SW 1.0 and 2.0 part of the vision team.
- Light-weighted edge device containing ads player and inference module by optimization.
- Proposed a custom DAG Scheduler pipeline for asynchronous ADs update.
- Implemented a data processor for raw edge logs to provide realtime aggregation.
- Implemented people counter module based on people tracking module, used in COEX convention hall.
LUXROBO | South Korea (Jan 2020 - Jul 2022)
Machine Learning Engineer (전문연구요원 편입)
- Implemented MODI Python API, PyMODI for MODI AI KIT.
- Enhaned usability of MODI by implementing SWs including MODI Firmware Updater, VirtualMODI.
- Led AI Team to win 2nd place in AI Championship 2020.
- Conducted a side-research on depth completion, published our research to CVPR 2022.
- Structured an automated data pipeline recommendation system APIs of LMS.
- Received the best peer (of AI) prize in 2021, nominated by colleagues.
Schlumberger UK | United Kingdom (Mar 2018 - Sep 2018)
Machine Learning Engineering Intern
- Implemented a machine learning program which predicts an optimal node group for small-sized reservoir simulation in HPC environment.
NEOWIZ | South Korea (Aug 2017 - Sep 2017)
Machine Learning Engineering Intern
- Implemented a deep musical note generator for a rhythm game called Tapsonic.
Research Experience
GuideFormer: Transformers for Image Guided Depth Completion
Kyeongha Rho*, Jinsung Ha*, Youngjung Kim
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2022
[pdf]
- We propose a fully transformer-based architecture for depth completion.
User-Guided Colorization Using Optimization and Learned Similarity
Jinsung Ha*, Shuyu Lin*, Ronald Clark
- We propose an optimization based deep learning approach for user-guided colorization.
Thesis
Self-Supervised Learning for Depth Image Completion and Enhancement
Co-Supervised by Dr Ronald Clark and Dr Sajad Saeedi at Dyson Robotics Lab
Achieved First Class Honours (Distinction)
- The objective is to use deep learning to enhance the depth data thus it can be used to its full benefit in robotic perception tasks.
Education
Imperial College London | London, United Kingdom (Oct 2014 - Oct 2019)
MEng Computing (Artificial Intelligence), Achieved 2:1
학석사통합과정 전산학 (인공지능), GPA 3.3/4.0
- Year 1. Programming, Databases, Architecture, Hardware, Discrete Maths, Maths Methods, Logic, Reasoning about Programs
- Year 2. Artificial Intelligence, Algorithms, Operating Systems, Networks, Software Designs, Statistics, Concurrency, Compilers, Models of Computation, Computational Techniques
- Year 3. Machine Learning, Computer Vision, Robotics, Web Security, Distributed Algorithms, Graphics, Advanced Databases
- Year 4. Deep Learning, Reinforcement Learning, Natural Language Processing, ML for imaging, Maths for ML, Machine Arguing, Distributed Ledgers, Software Engineering for Industry
Projects
Awards & Activities
창업진흥원장상 (2nd Place in LG Science Park Section) | AI Championship 2020
LUXROBO AI Team
- Developed a deep ensemble network to detect anomaly in noise inspection of electronics.
Opensource Contributon 2021 | Open Source Software KR
Participant
- Contributed to pytorch-tutorials-kr by translating official docs in Korean.
Certifications
| Name | Issued by | Expires on |
|---|---|---|
| Tensorflow Developer Certificate | Tensorflow | Jan 2024 |
| Terraform Associate (003) | HashiCorp | Sep 2025 |
Opensource Contributions
| Project | Role | PR # |
|---|---|---|
| pymodi | Maintainer | >= 97 |
| modi-firmware-updater | Author | >= 1 |
| iamport-rest-client-python | Maintainer | >= 43 |
| ray | Contributor | 38496 |
| pyupbit | Contributor | 5, 7, 8, 9 |
| pytorch-tutorials-kr | Contributor | 276 |
| backend-interview-question | Contributor | 24 |
| virtual-modi | Author | >= 1 |
Languages
Korean(native) and English(fluent)
references available upon request