Capstone · Team — sole AI dev · In production

An AI CRM built into the live social-matching service Inpose — it automatically clusters feed images by visual semantics and learns each user's behavior to power personalized job-posting and feed recommendations.

My role — Sole AI developer on the team: designed and built the entire ML pipeline single-handedly — embedding extraction, clustering, inference server, user-affinity module, and production deployment.

• Extracted CLIP ViT-L/14 embeddings from 384,761 images → PCA + K-Means (K=500); chose K-Means over DBSCAN/HDBSCAN after analyzing failure modes on high-dimensional embeddings
• Real-time inference at ~13ms/image (~78.8 img/s, GPU) via cosine-distance centroid matching
• Behavior-weighted user affinity module (view/like/comment/scrap), deployed with FastAPI

Details →

Grand Prize ×2 · Team Lead

A service that recognizes bulky household waste from a camera in real time and automatically calculates the disposal fee per item. Won top prize at two competitions.

My role — Team lead (4-person team). Owned the full codebase end-to-end and led the research report — from data pipeline to model training, evaluation, and the service prototype.

• YOLOv8n (imgsz=800, 300 epochs) → mAP@50 98.1% · mAP@50-95 93.2%
• Built the full data pipeline on the NIA dataset: JSON→YOLO label conversion, polygon/box parsing, multi-class augmentation (albumentations)
• Designed an end-to-end service prototype with automated fee calculation

Details →

Team project · Konkuk Dream Semester

A defense-reconnaissance AI that detects camouflaged/occluded military targets (personnel, vehicles, APCs). Team project under Konkuk University's Dream Semester program.

My role — Data collection & annotation, plus the modeling: custom CBAM architecture, Focal-Loss tuning, and the ensemble/SAHI evaluation pipeline.

• Built the dataset: web scraping + synthetic ARMA 3 scenarios, 1,800+ images labeled via Roboflow (3 classes)
• Designed a custom YOLOv8m + CBAM + Focal Loss model → mAP@50 0.924 vs 0.883 baseline (+4.1%p; mAP@50-95 0.698 vs 0.639)
• Built a multi-scale WBF ensemble + TTA and a SAHI sliced-inference pipeline for small/distant targets

Details →

Team (AI role) · Agentic dev

An AI-assisted traffic-control system that pre-screens road snapshots for animals/carcasses so operators only review images that need a human decision. Built through a controlled agentic-AI workflow.

My role — Owned the AI pipeline: YOLOv8 training on 67,275 images (3 classes), a data-existence verification gate, a two-stage confidence threshold (0.3 detect / 0.6 alert), and the inference→webhook alert pipeline.

• Validation mAP@50 0.994; honest external test (12 imgs) → precision 1.0, recall 0.67
• Rejected the LLM's single-threshold design for a two-stage detect/alert split tuned on held-out data

Details →

Solo · Classical CV (no DL)

Two vision tasks solved with classical image processing only — a counterpoint showing I understand the fundamentals beneath modern detectors.

My role — Built both pipelines solo.

• Eye Image Segmentation — pupil-center detection via threshold → morphology → multi-radius circular template matching (IR/RGB/Depth)
• Pathology Slide Classification — normal vs cancer-suspected patches via hand-engineered features (tissue ratio, morphological irregularity, local density) + rule classifier

Details →

Live Service · PM + Frontend · Team of 4

The official portal of Konkuk University's Student Welfare Committee, digitalizing campus rental and printing operations. Currently live in production.

My role — Project planning (PM), frontend architecture & implementation, deployment, and ongoing maintenance/operations.

• Digitalized rental & printing workflows, drastically reducing administrative overhead

Details →

Grand Prize · Frontend · Team

An all-in-one recruitment web app that handles application collection, pass/fail management, bulk SMS, and interview scheduling in one place. Grand Prize at Kuit 6th Demoday.

My role — Frontend developer: built the Home / My / Manage / Notification pages and implemented & integrated the smart scheduling feature.

• Smart interview scheduling with automated time-slot recommendation and applicant self-adjustment

Details →

Published · KICS 2026

DNN-based resource allocation for progressive image transmission over multi-frequency-band MIMO interference channels — jointly optimizing per-packet spectral efficiency, spatial multiplexing, sub-band combination, and transmit power via a PSNR-based, end-to-end differentiable loss.

Details →

Industry-Academia Research

Resource allocation for D2D wireless networks. Frequency-allocation comparison of five methods (Genetic, Game Theory, DNN, Hungarian, Graph Coloring) vs Full Search — Genetic reached 99.9%. Plus a joint power–frequency study — a self-designed DNN reached 97.6% of Full Search, with GNN variants (ResourceAllocationGNN, JCPGNN) explored against an FS+WMMSE baseline.

Details →

Neuromorphic · SNN

A spiking network that learns a digit → letter association purely through spike-timing-dependent plasticity (STDP) — LIF neurons + step-conductance synapses, taught with nontarget inhibition, recalling the letter from the digit alone.

Details →