Projects
Oil & gas company · via Accenture · 2026
Satellite Methane Detection & Agentic Incident Investigation
Leading the Data & AI stream of a methane detection platform going to production in October 2026. It detects and sizes plumes at industrial sites from public satellite imagery, cross-checks them against ground gas sensors and maintenance records, and has a Deep Agent find the root cause, recommend next-best actions, and write a cited report for human review.
Accenture · production SaaS, Germany-based team (remote) · 2025–2026
AI Agents for a Production GenAI Ad Creation Platform
Shipped AI agents, an LLM evaluation gate, and RAG ingestion to the Python/FastAPI AI orchestration service of a multi-tenant GenAI platform used in production by global consumer brands to brief campaigns, extract brand guidelines, and generate on-brand ad creatives.
Telecommunications operator · via Accenture · 2026–present
Self-Hosted Real-Time Voice AI Agents
Building a self-hosted voice agent platform (speech-to-text, LLM, text-to-speech) that lets callers talk to AI customer-service agents in Indonesian and other languages over the web, a mobile app, and phone lines, with every model running on private GPUs. In progress.
Telecommunications operator · via Accenture · 2024–2025
Retail Store Network Optimization with Geospatial ML
Built a geospatial decision engine that recommends, for every customer-service store in a national network, whether to keep, resize, relocate, or close it and where self-service kiosks should go, with an LLM writing the justification for each store.
Beauty & personal-care company · via Accenture · 2024
GenAI Video Analysis for Affiliate Creators
Contributed to a GenAI video-understanding solution that analyzes affiliate creators’ short-form videos in minutes instead of hours of manual review: Gemini scores each video scene by scene and tells brand teams why it sells and what the creator should do next.
FMCG company · via Delman Data Lab · 2022–2024
Automated Daily Sales & Stock Reporting
Automated the daily sales and stock reports an FMCG company builds from its modern-trade retail partners’ files, cutting operational costs by 50% and daily processing time by 99%, with no human errors found after the switch.
Personal research project · open source · 2026
PlanSpot: One-Shot Symbol Spotting on CAD Drawings
Box one symbol on a floor plan and PlanSpot finds every other instance in any orientation, with no training for that symbol. It compares frozen DINOv2 feature matching with a classical pixel-matching baseline on real CAD drawings, with drawing-level splits and per-query cut-offs. A second phase fine-tunes DINOv2 with LoRA to segment 1–3 px wall strokes.
Consumer goods company · via Accenture · 2024
Customer Data Cleansing & RFM Segmentation at Scale
Built containerized data-cleansing and RFM segmentation pipelines covering 5.8M customers across brands and online marketplaces, with 77 K-means models and an LLM labeling each cluster in business terms.
Energy group · via Accenture · 2025
Responsible AI Governance & AI Technology Assessment
Assessed AI use cases across a multi-business-unit energy group, produced a technology selection guide, and ran hands-on Responsible AI risk assessments with explainability, LLM-quality, and PII tests.
Manufacturing group · via Delman Data Lab · 2022–2024
Procurement Master Data & Price Analytics
Unified item, supplier, and transaction data from three business units of a manufacturing group into golden records, so procurement could compare prices across suppliers. The comparison showed up to 50% potential savings from choosing the best-priced supplier.
Daewoong Pharmaceutical · 2021–2022
Real-Time Encapsulation Quality Detection
Built a computer vision model that judges droplet encapsulation quality from sequences of camera frames (93.5% test accuracy) and predicts the machine’s frequency setting, delivered as a standalone real-time app beside the encapsulation machine.
Daewoong Pharmaceutical · 2021–2022
Sano: Personal Health Assistant Chatbot
A Telegram health assistant that reminds patients to take their medication, identifies drugs from text or a photo, recommends similar drugs when one is out of stock, and points to a nearby pharmacy.
Daewoong Pharmaceutical · 2021–2022
Customer Review Analytics for Online Stores
Analyzed Korean customer reviews of online-store products: a BERT model separates positive from negative reviews, a BART model summarizes each side per product, and word clouds surface the most frequent terms, all in a Flask + React dashboard.
Pusan National University Research · 2018–2024
Q-Learning for 6TiSCH Network Formation and Transmission
Ph.D. research that makes industrial IoT (6TiSCH) networks learn: Q-Trickle speeds up network formation, and ACI-RPL makes routing more reliable with cooperative multi-agent Q-learning. Both were validated in simulation and on up to 32 real nodes.
MUFKO, South Korea · 2019
MUFKO Halal App
Backend developer for MUFKO (Muslim Friendly Korea), an app that scans Korean snacks for halal-relevant ingredients and lists halal restaurants and prayer places. It has 100K+ downloads and a 4.6-star rating on Google Play.
Center for IoT Innovation (CITI), NTUST, Taiwan · 2018
Human Activity Recognition for Security Check & Hand Washing
A video system that checks whether people follow a security-check routine and a hand-washing routine. It extracts OpenPose keypoints from each frame and classifies 20-frame sequences with an LSTM, reaching 98% test accuracy on security check and 63% on hand washing. It continued my bachelor-thesis topic, human activity recognition on video.
Skills
Agentic AI
Building AI agents and multi-agent systems with tools, MCP, and evaluation gates, from proof of concept to production.
- Pydantic AI — production deep agents with typed output
- Deep Agents — planning agents with tools and skills
- LangGraph — stateful agent orchestration
- LangChain — LLM application building blocks
- MCP — Model Context Protocol tools
- LLM Evaluation — pydantic-evals, LLM-as-judge, CI gates
- Langfuse — LLM tracing and eval datasets
- Agent Harness — spec-driven autonomous delivery
GenAI & LLMs
Applying and serving large language models through RAG, prompt engineering, self-hosted inference, and API integration for production use cases.
- RAG — pgvector and BM25 hybrid search
- Prompt Engineering — CoT, ToT, ReAct, Reflexion
- vLLM & TensorRT-LLM — self-hosted LLM and STT serving on GPUs
- Qwen — open-weight LLMs (AWQ quantized)
- Claude API — Anthropic SDK and tool use
- OpenAI / Azure OpenAI — GPT and image models
- Gemini / Vertex AI — Google multimodal models
- OpenRouter — multi-provider model routing
Computer Vision
Image classification, object detection, segmentation, and real-time inspection systems using deep learning.
- PyTorch — CNN, ViT, custom architectures
- OpenCV — image processing and pipelines
- Object Detection — YOLO, Faster R-CNN
- Semantic Segmentation — U-Net, DeepLabV3+ with ConvNeXt
- Self-Supervised Learning — SimCLR pretraining, timm backbones
- Image Layer Decomposition — LayerD, LaMa inpainting, OCR
- OCR — PaddleOCR, Docling
- Real-time Inference — production-line inspection systems
NLP & Voice AI
Text understanding with transformers, and real-time voice agents that listen, think, and speak with low latency.
- Transformers — BERT, RoBERTa, T5 family
- HuggingFace — model hub and fine-tuning
- Text Classification — multi-label and multi-class
- Information Extraction — NER, relation extraction
- Pipecat — streaming voice agent pipelines
- Whisper (STT) — speech-to-text on TensorRT-LLM
- GPT-SoVITS (TTS) — Indonesian TTS fine-tuning with a G2P frontend
- Silero VAD & WebRTC — turn-taking, barge-in, real-time transport
Remote Sensing
Analysis of satellite and aerial imagery for land cover classification, change detection, and geospatial feature extraction.
- Land Cover Classification — from multi-spectral imagery
- Change Detection — temporal satellite image analysis
- Feature Extraction — infrastructure and asset mapping
- GIS Integration — geospatial data pipelines
- Sentinel-2, Landsat, EMIT — multispectral and hyperspectral imagery
- Apache Sedona & H3 — spatial SQL and hexagonal grids at scale
- GeoPandas & rasterio — vector and raster processing
ML & Deep Learning
Classical machine learning and deep learning for predictive modeling, optimization, and data-driven decision systems.
- Scikit-learn — classical ML and pipelines
- XGBoost — demand and tabular models
- K-means Clustering — customer and visitor segmentation
- Optimization — integer programming, genetic algorithms
- PyTorch — neural network training
- Keras — rapid prototyping
- Reinforcement Learning — Q-Learning, policy gradient
- Feature Engineering — tabular and time-series data
Data, MLOps & Cloud
Building data pipelines, ML platforms, APIs, and cloud infrastructure that take models to production.
- Databricks — MLflow, Unity Catalog, GPU jobs
- PySpark & Airflow — medallion pipelines and orchestration
- Kubeflow — ML pipelines
- AWS — ECS, EKS, S3, ECR, CloudWatch
- Terraform & Docker — infrastructure as code and containers
- GitHub Actions — CI, test and eval gates
- FastAPI — async REST APIs
- GraphQL & OAuth 2.0 — API backends with Google login
- Flask — lightweight web services
- PostgreSQL — relational data modeling
- ETL Pipelines — data ingestion and transformation
- Pandas & NumPy — data wrangling and analysis
- Streamlit — data apps and dashboards