Projects

Anonymous voice support app for people preparing for jobs or major exams, or pushing through burnout and depression. No follows, no networking, no influencer dynamics.

Mobile invoicing for field technicians, dog walkers, and private tutors. Create invoices on-site with voice input or templates, attach proof photos, and share with clients immediately.

Led the data pipeline and audit-risk labeling workflow for global news intelligence.

• Built a data ETL pipeline for news metadata and source-credibility labeling across 17 countries, partnering with regional Deloitte SMEs to map trusted sources
• Built relevance + audit-risk classification using title/summary/company tags; curated a goldenset with SMEs for evaluation
• Implemented client-specific processing and DUNS-based entity updates to handle M&A changes near real time; processes 40–50K articles/day and ~20K labeled records
• Delivered RiskSensing API on AKS with QA test suites, OpenTelemetry monitoring, and LLM cost optimization from ~$40/day to ~$6/day (GPT-4/4o/5)

Developed a cross-platform desktop notification system for Claude Code that enhances developer productivity by providing real-time alerts when AI tasks complete, errors occur, or user input is needed. Features native desktop notifications across macOS, Windows, and Linux with customizable sound alerts, global and project-specific settings, and seamless integration with Claude Code workflows.

Led the end-to-end development of an intelligent nursing diagnostic system. I designed and implemented a Retrieval-Augmented Generation (RAG) system that references a knowledge base of 80+ documented nursing scenarios. When new patient data is entered, the system retrieves the top 3 most similar scenarios to inform its diagnostic suggestions.

I was responsible for the entire audio processing pipeline. My primary role was to extract and analyze audio from raw video footage, tackling the significant challenge of low-quality audio in Korean. I developed a noise reduction process using spectral subtraction and a filtering logic to isolate the child's voice from background noise and parental speech, significantly improving the quality of data for the model.

Developed a private LLM with RAG using LangChain and vector databases (FAISS, Qdrant) to support Q&A, summarization, and enterprise document retrieval. Built an auto-updating vector index that detects document changes in real time and compared LLMs (Llama-2, Falcon, GPT4ALL) for accuracy and latency.

Developed a multimodal AI system enabling the Pepper robot to navigate autonomously. The robot uses Microsoft HoloLens for real-time environment scanning, obstacle detection, and spatial mapping.

I took over a stalled project that used a traditional NLP model and a Unity 3D avatar. I completely redesigned the system by integrating the GPT API for fluid conversation and OpenAI's Whisper API for robust speech-to-text and text-to-speech capabilities. The virtual avatar was replaced with a physical Pepper robot for tangible user interaction.

Developed a novel methodology combining Finite Element Analysis (FEA) with Machine Learning to predict structural integrity in aerospace components.