Open to opportunities

I'm ChaninAI/LLM Developer

I build enterprise AI systems that real people use every day — not prototypes. Engineering background, hands-on LLM ops.

100+
Users served
2+
Years in AI
2
Research papers
2
Production systems
Chanin Jung
LLM / RAG·AWS·Python·Dify · n8n·Cloudflare Pages·Langfuse·YOLOv5 / GAN·XGBoost· LLM / RAG·AWS·Python·Dify · n8n·Cloudflare Pages·Langfuse·YOLOv5 / GAN·XGBoost·
Work

What I've Built

Enterprise LLM agent answering a construction-regulation question inside Mattermost — English UI, listing the building and school-facility regulations to review
Production

Enterprise LLM Agent Platform

AI agent serving 100+ users across 6 departments at a 1,700-employee construction engineering company. Users ask regulatory questions in Mattermost — the agent retrieves answers with cited legal references.

PythonDifyn8nAWSTerraformKeycloak
Key points
  • Dify-based LLM Agent autonomously calls 3 data sources: Korean Construction Standards API, National Legislation API, Gemini web grounding
  • Full round-trip pipeline: Mattermost → n8n → Dify Agent → n8n → Mattermost
  • 21 pods, 6 open-source services on AWS — Keycloak SSO across 5 tools, CI/CD via GitHub Actions
  • Regulatory document search time: 1 week → 1 minute
  • Query cost: ₩33/query (~$0.025)
Huko News homepage — Daily Korean News
Production

Huko News — A Live Full-Stack Web Service

A live Korean-learning news website I designed, built, and run end-to-end — a real product, not a prototype. I own the whole stack: front-end, content engine, and deployment. hukonews.com ↗

HTML/CSS/JSPythonLLMCloudflare PagesGitHub ActionsSEO
Key points
  • Designed & built the site from scratch — responsive front-end (HTML/CSS/JS), article / category / search pages, and a lesson-material library
  • Engineered the Python content engine: news collection → LLM rewriting & level-grading → trilingual translation (KOR/ENG/HU) → image sourcing → TTS audio
  • Ships daily via GitHub Actions to Cloudflare Pages (07:00 KST), with SEO automation (sitemap, IndexNow) and AdSense monetization
POSCO Academy analysis-results slide: defect frequency by discharge-stage-2 temperature, a table ranking key-factor candidates across four methods, and the optimal 25.6–29.7°C window
POSCO Academy · Award

Li-ion Battery Defect Rate Reduction via XGBoost

Analyzed real production data from a lithium-ion battery formation process to identify defect-causing factors and build a predictive model to reduce the defect rate.

XGBoostRandom ForestANOVAPythonPandas
Key points
  • Dataset: 44,078 observations × 113 process variables + 26 time-series variables
  • Feature selection via ANOVA, Chi-squared, Decision Tree — identified charge/discharge equipment and temperature as key factors
  • Defect rate reduced from 3.12% → 0.45% using XGBoost prediction model
  • Domain insight: outlier temperatures (25°–30°C) validated against literature and preserved, not removed
YOLOv5 plant-disease detection results: a model comparison table with YOLOv5s leading at mAP50 0.991, the F1-confidence curve and the confusion matrix
GAN-generated tomato ripeness images, animating from noise to a realistic tomato
GAN samples
POSCO Academy · Award

Smart Farm AI — Tomato Ripeness & Disease Detection

Built a real-time YOLOv5-based detection system for tomato ripeness (3 classes) and disease (10 classes). Resolved class imbalance by training a GAN from scratch to generate synthetic ripeness images — and published the augmented dataset publicly on Roboflow.

YOLOv5GANComputer VisionFlaskArduinoPython
Key points
  • Ripeness dataset: 2,448 images (turning/pink/red — class imbalance in "turning")
  • GAN generated 600 synthetic "turning" images → final training set: 3,088 images; FID-optimized parameters
  • Turning class mAP50: 0.969 → 0.991 after GAN augmentation; final ripeness mAP50: 0.973
  • Disease model: compared YOLOv5/v6/v8 → YOLOv5 selected at mAP50 0.991
  • RFID + Arduino for tomato location tracking; Flask web dashboard for real-time detection feed
Research

Publications & Academic Work

i3CE 2026 — ASCE 2026 · First Author

BM25-based Retrieval-Augmented Compliance Verification for Korean Construction Standards  ↗ ASCE Library

Proceedings 2026 — not published yet; the ASCE volume is due next year. The link opens the i3CE proceedings listing on the ASCE Library.

BM25-based RAG pipeline architecture

A domain-specific RAG system for Korean construction regulations. A structure-preserving Markdown–HTML preprocessing pipeline reduces token overhead by 78%, and BM25-based sparse retrieval with morphological decomposition enables clause-level terminology alignment and citation-grounded verification. Evaluation on 50 expert-curated compliance questions shows improved accuracy (76.0% → 84.0%) and elimination of false positives (11 → 0). Presented at ASCE i3CE 2026, Songdo, June 2026.

84.0%
Accuracy (vs 76% standalone)
82.0%
Recall@5
0
False Positives (vs 11)
78%
Token reduction
BM25RAGCompliance VerificationConstruction StandardsPython
IEEE Access — SCI 2023 · First Author

Counter-Rotating Hoop Stabilizer and SVR Control for Two-Wheels Vehicle Applications  ↗ IEEE

Lab demo — the prototype holds balance at a 23° lean angle (loops)

The capstone design project won the Grand Prize (2021) and was awarded research funding to continue development. The PID-controlled reaction wheel system was then fully redesigned for this SCI paper: replaced PID with SVR (Support Vector Regression), and replaced the single reaction wheel with a counter-rotating hoop — two wheels spinning in opposite directions to achieve more stable, bidirectional balance control. SVR predicted the required PWM voltage at each tilt angle; hyperparameter optimization with RBF kernel achieved 98.2% correlation.

100%
Balance correction accuracy
±23°
Tilt correction range
2.13 s
Avg. recovery time
98.2%
SVR correlation (RBF kernel)
SVRPID ControlReaction WheelEmbedded SystemsArduinoMATLAB/Simulink
Experience

Career

Feb 2024 — Apr 2026

AI Developer — LLM / RAG Engineer

Dongmyeong Engineering & Construction (DMEC) — Seoul, Korea
  • Built and operate enterprise LLM agent platform serving 100+ users on AWS
  • Production RAG pipelines with Elasticsearch for regulatory / legal document search
  • AI agent & chatbot development with Dify; workflow automation with n8n
  • LLM observability and cost monitoring with Langfuse
  • Infrastructure: 6 services, 21 pods; Keycloak SSO; Terraform; GitHub Actions CI/CD
Oct 2014 — Feb 2019  (4 yrs 3 mos)

HVAC & Mechanical Design Engineer

Building Systems Design Firms — Seoul, Korea
  • HVAC system design and CAD documentation for large-scale construction projects
  • Deep familiarity with Korean construction standards, legal codes, and regulatory documents
  • This domain knowledge became the foundation for the AI systems I later built
Education

Academic Background

Degrees
2026

M.S. Computer Science

2026 — Present
+ Specialized in Artificial Intelligence
Eötvös Loránd University — Budapest, Hungary

Stipendium Hungaricum Scholarship Recipient: Awarded full government scholarship for Master's degree in Artificial Intelligence (Budapest, Hungary).

2023

B.S. Mechanical & Biomedical Engineering

Graduated 2023
+ Double Major in Philosophy
Ewha Womans University — Seoul, Korea

Engineering thesis (Grand Prize): Reaction Wheel balance & direction control prototype using PID control — awarded research funding, extended into first-author SCI paper (IEEE Access, 2023).

Intensive Programs
2025

AIFFEL LLM Engineering Course

Aug 2025 — Nov 2025
Modulabs — Korea

204 hours · RAG & LLM Agent development

2023

POSCO AI & BigData Academy

Oct 2023 — Dec 2023
POSCO HRD × POSTECH — Korea

Intensive ML/DL program · Project Excellence Award

Awards & Certifications

Recognition

Award
Capstone Design Grand Prize
Ewha Womans University
Award
Project Excellence Award
POSCO AI BigData Academy
Certification
ADsP — Data Analysis Semi-Professional
Korea Data Industry Promotion Institute
Certification
SQLD — SQL Developer
Korea Data Industry Promotion Institute
Tech Stack

Tools & Technologies

LLM / AI
DifyLangChainLangfuseRAGPrompt Engineering
Search & Retrieval
ElasticsearchBM25Vector DB
Infrastructure
AWSTerraformKubernetesDocker
Backend & Automation
PythonFastAPIn8n
Auth & Monitoring
KeycloakOAuthLangfuseCloudWatch
Languages
Korean (Native)English (Business Level)
About

A little about me

Engineer-turned-AI-developer who ships real systems — from front-end to infra.

I started in HVAC & mechanical design, moved into AI/LLM engineering, and now build enterprise RAG + agent systems that people rely on daily. I like owning the whole path: designing it, building it, deploying it, and keeping it running.

Currently pursuing an M.S. in Computer Science (AI) at Eötvös Loránd University, Budapest.

Contact

Let's build
something.