Hyungmo

Kim

SW/ML Engineering

Senior ML Software Engineer

ㅡ

Summary

I am a polyglot programmer for any problems that can be framed, designed, and interpreted as a program; applications, platforms, models, processes, strategies and team topology.

I mostly focus on learning immutable values like abstract structures, first principles, and mental models that can be generically applied to practical topics.

ㅡ

Experiences

Multi-task on-device model / ML Software Engineer

2025 - 2026, Hyperconnect LLC / MatchGroup AI

• Design and propose to train multi-task vision on-device model to replace previous models in iOS/Android model pipeline based on literature survey
• Propose modular model architecture and lead machine learning engineers in different timezones to train different parts concurrently without blocking each other
• Distill pre-trained CLIP (ViT) models into MobileNet variant to support better generalizability and semantic search accuracy
• Reduced the number of model invocation counts to match product requirements on end-to-end latency (x3 faster)

On-device ML platform as a service / Acting Manager

2023 - 2025, Hyperconnect LLC / MatchGroup AI

• Design and propose a prototype of the platform for easy integration of mobile inference capabilities
• iOS demo application development and Android demo application multithreading for third-party PoC delivery
• Develop integrated Python script for conversion, encryption and preprocessing insertion between various formats such as TFLite / TorchScript / ONNX / CoreML
• Develop a generalized benchmark application for testing in mobile environments
• Lead software engineers and machine learning engineers for several projects on server and on-device model optimizations
• Supervise tech blog writing: https://hyperconnect.github.io/2026/01/23/On-device-Face-Verification-Pipeline-Optimization.html

GDPR compliance for model training data / Acting Manager

2024, Hyperconnect LLC

• Select points where problems may occur across the entire learning infrastructure, including in-house cloud and on-premise, and take preemptive action
• Suggest problem specification and software design for automation of each part, supervise team members

Serving 70B LLM on-premise / Acting Manager

2024, Hyperconnect LLC

• Conduct data research and benchmark to cost-effectively serve 70B LLM in real time
• Propose and build a system that provides services at no additional cost using already purchased on-premise in 2 weeks
• Define and breakdown a complex problem into delegatable and deliverable pieces
• Cost savings of hundreds of millions of won per month

Unstructured training data ETL pipeline / ML Software Engineer

2022 - 2023, Hyperconnect LLC

• Replace existing legacy data pipeline, which are vulnerable to domain/schema/business rule changes and lack functionalities such as backfill/catalog, in accordance with the relevant product transfer schedule
• There were no existing domain terms (ubiquitous language) and automated systems related to data/labels, so I’ve proposed a new one
• Automation of various media processing such as image/audio/text/video

Real-time on-device audio classification runtime
/ ML Software Engineer

2021, Hyperconnect LLC

• Development of a TFLite-based executor that runs a context-based model that improves the limitations of existing keyword detection models on Android/iOS devices
• Advanced Python tools for TFLite model conversion, metadata operations
• Reimplementation of inference engine code base for software quality

CUDA-based LLM inference runtime / ML Software Engineer

2021, Hyperconnect LLC

• Additional CUDA kernel implementation of NVIDIA’s FasterTransformer, bug fixes, gpu memory optimization, various heuristic implementations
  (checked correct behaviour up to 13B)
• Run actual services as a backend on NVIDIA Triton inference servers
  (4B scale)
• Up to 50x improvement over RPS of existing technology (at that time)

On-premises GPU cluster for research
/ ML Software Engineer

2020 - 2024, Hyperconnect LLC

• Review of 50 PF GPU cluster system design and technical specifications based on NVIDIA SuperPod architecture
• Achieving more than twice the cost efficiency compared to AWS Cloud
• 400TB distributed storage configuration and GitOps-based management system design
• Constructing a deep learning research environment that allows large-scale distributed learning through slurm scheduler using Ansible and systemd

Real-time on-device image classification engine
/ ML Software Engineer

2020, Hyperconnect LLC

• Development of a TFLite-based inference engine SDK that runs a new lightweight model with faster inference time than existing image classification/segmentation models on Android/iOS devices
• TensorFlow 2 Keras quantization related technology PoC in progress
• Application of TFLite GPU/XNNPack delegate technology for hardware acceleration
• Development of Android/iOS demo app for testing in WebRTC environment

K-supercomputer, Chundoong / System Administrator

2017 - 2020, SNU Thunder Research Group

• Water-cooled/oil-cooled cluster system management of 200 AMD+NVIDIA heterogeneous GPUs and Infiniband interconnect network
• Service for over 300 users for educational/research purposes

Deep learning framework for Samsung NPU (SNPU) system
/ Researcher

2019, Samsung Electronics

• Development of CNN benchmark to analyze Samsung Neural Processing Unit performance
• Implement distributed learning benchmarks of 4 types of CNN (VGG, ResNet, DenseNet, Inception) with cuDNN+MPI

HPC testbed system and multi-node GPU benchmarks / Researcher

2017 - 2019, Ministry of Science and ICT

• Build and manage 32 NVIDIA Tesla V100 physical testbed systems
• Development of OpenCL and CUDA versions of GPU benchmark
  (SNU NPB 2019)

ㅡ

Education

Master’s degree in CSE / Seoul Nat’l Univ.

2017-03 - 2020-02

• SNU Thunder research group (http://aces.snu.ac.kr/)

Bachelor’s degree in CSE / Seoul Nat’l Univ.

2013-03 - 2017-02

ㅡ

Awards

2017 National Supercomputing Contest Grand Prize
/ UNIST President’s Award

2nd prize, learn more: https://eng.snu.ac.kr/snu/bbs/BMSR00005

ㅡ

Publications

International Papers

1. SNU-NPB 2019: Parallelizing and Optimizing NPB in OpenCL and CUDA for Modern GPUs

Youngdong Do, Hyungmo Kim, Pyeongseok Oh, Daeyoung Park, Jaejin Lee

IISWC '19: Proceedings of the 2019 IEEE International Symposium on Workload Characterization, Orlando, FL, USA, November 2019, DOI: https://doi.org/10.1109/IISWC47752.2019.9041954

ㅡ

Skills

I leave the traditional “Skills” section empty, as specific programming languages, frameworks, or tools feel less significant today. Instead, I prefer to highlight what I can actually accomplish.

Optimization

• I’ve started my career from optimizing numerical analysis programs in C using GPGPU for benchmarking supercomputers. My first project in Hyperconnect was to migrate the underlying technology of our on-device inference SDK from TF1 to TF2. I understand low-level mechanisms of how computers utilize their resources and how neural networks are implemented as software and mapped to hardware.
• I have intuitions and experiences on how to identify bottlenecks in ML systems and resolve them step-by-step.

Engineering

• I am comfortable reading and writing across programming languages in multiple programming paradigms, even unfamiliar ones.
• I can comprehend cross-functional domains in the AI engineering era and translate them into well-defined SW or ML problems. I can generate concrete design and execution plans on them considering resources and uncertainties.

Management

• I have a basic understanding of people management that I’ve experienced while working as an acting manager in two different organizations; MLOps team and central AI team.
• I’m an experienced agile practitioner; I’ve managed several projects with different cross-functional members and methodologies using issue tracking and communication tools. I can identify blockers, collaboration areas and bad smells on different technical domains.