MSc thesis
Gated YOLO
Learned channel gates that prune YOLOv6 for edge deployment: 12% fewer MACs at 1.3 mAP below the unpruned detector on VOC, with TensorRT FP16 and QAT INT8.
Code →Daegu, South Korea
ML Engineer, Computer Vision & Edge AI
I build computer vision models and ship them to the edge. I lead the AICU Gait Scanner, from pose and gait models to optimized on-device builds for iOS and Android, and the services behind them.
8+ years of software engineering, 3+ years of applied ML. MSc Computer Science, Kyungpook National University.

Experience
Machine Learning Researcher, Gait Scanner lead
Software Developer, part-time contractor
Software Engineer Specialist
Co-founder, lead software developer
Research
MSc thesis
Learned channel gates that prune YOLOv6 for edge deployment: 12% fewer MACs at 1.3 mAP below the unpruned detector on VOC, with TensorRT FP16 and QAT INT8.
Code →JICCE 2023, co-author
Edge AI detection for fixed cameras on embedded hardware.
Paper →Dataset, 2023
13,221 images and 166,192 labeled objects from fixed security cameras over highways and roads, in varied weather and light.
Dataset →Projects
Everything here runs on my own k3s cluster: Helm charts, GPU inference with Ollama, VictoriaMetrics and Grafana alerting.
Live car auctions with ML pricing (LightGBM, conformal intervals), CLIP photo checks in int8 ONNX and weekly retraining behind a promotion gate.
Self-hosted finance app with an MCP endpoint, so an AI agent can manage it through task-level tools behind a token and a NetworkPolicy.
Document line segmentation built on CRAFT word detection: deskewing, polygon grouping into lines, and a web API.
Speaks the OpenAI Responses, Chat Completions and Anthropic Messages APIs and runs them on local Ollama models. Tested at 95% coverage, signed images with SBOMs.
Hand gesture classifier trained from scratch, served as a playable game.
Writing
Hands-on posts on gait analysis, computer vision and AI tooling, written for AICU.
Markerless vs marker-based capture, the five steps from video to gait measures, and what studies show about accuracy.
gait.aicu.lifeReference systems, ICC, Bland–Altman limits of agreement and meaningful change, applied to published camera and phone tools.
gait.aicu.lifeThe five kinds of gait analysis system, from optical motion capture to markerless video and phone apps, and what each measures.
gait.aicu.lifeThe gait cycle, stance and swing, double support, cadence, step and stride length and symmetry, with normal values.
gait.aicu.lifeMatching YOLO11 person crops with torchreid OSNet, exporting to ONNX and comparing four variants on CPU.
blog.aicu.lifeYOLO11 detection exposed as an MCP tool over stdio, with protocol latency, cold start and token cost measured.
blog.aicu.lifeA 143-line gateway with per-role tool allow-lists and an audit log: 57–69% fewer tool tokens for 0.6 ms per call.
blog.aicu.lifeA person detector on every camera frame at 30 FPS: YUV to RGB in C++, a GPU to NNAPI to CPU fallback, and Metal on iOS.
blog.aicu.lifeHU values of eight tissues measured on a public chest CT, plus three DICOM traps that silently corrupt them.
blog.aicu.lifeUsing NVIDIA Kimodo motion diffusion to synthesize rare gait patterns, and why its free foot-contact labels matter more.
blog.aicu.lifeToolbox
Kyungpook National University · MSc, Computer Science, Global Korea Scholarship (GKS) · Daegu, South Korea · 2024
INTEC · Software Engineering (Ingeniero de Software) · Santo Domingo, Dominican Republic · 2015 – 2019
The fastest way to reach me is email.