Beege!

Centralized system for organizing DICOM metadata, file references, and clinical study data — the first successful effort to get all the data in one place for annotation and ML training. Ideated the original system and led product management, delivering features including manufacturer aliasing, acquisition project tracking, and versioned ML splits for EU AI Act reproducibility. Centralized ingestion also uncovered and resolved data quality and duplication issues.

Massively concurrent pipeline for ingesting medical imaging data, preparing annotation datasets, and running bulk inference with in-house AI models. Conceived the original idea and built the proof of concept using Prefect before handing off to the team. Routinely processed hundreds of terabytes using spot instances and tuned autoscaling.

Internal tool for oncologists to label pathology whole slide images for cancer detection AI training, supporting distinct annotation modes — painting tissue areas, designating points, drawing cell contours, and sub-cellular labeling — on deeply zoomable WSIs. External contractor access with manager permissions enabled Lunit to scale annotation beyond its internal team.

Internal tool for radiologists to label medical images across chest X-rays, mammograms, and digital breast tomosynthesis, with queue-based and pool-based assignment and keyboard-shortcut-driven annotation. Took over as product manager and engineering manager. Introduced DBT modality support, multi-window current-prior comparison, cross-image annotation grouping, a reviewer feedback system, and an ML researcher API.

Remote management system for monitoring and controlling edge devices deployed at client sites. Built WebSocket-over-TLS communication between central and edge nodes, with relay node support for client network segmentation. All remote functions explicitly defined to pass security audits.

Backend for Robi, an AI-powered phone receptionist that routes callers to appropriate departments via Google DialogFlow for speech recognition, intent detection, and TTS. Architected the backend as a monorepo of six microservices controlled by feature flags.

Service enabling users to reserve and autonomously access rental cars without human intervention. Developed backend reservation and vehicle access APIs as part of an 8-service Spring Boot platform.

Team analysis examining whether League of Legends players had a meaningful win-rate change on wired vs. wireless connections. Contributed by querying petabytes of game and infrastructure data from Hadoop Hive; statistical analysis led by teammate. Findings delivered via Tableau dashboards.

Minute-resolution metrics collection system feeding game server health data to the Network Operations Center. Built collectors for server availability, login queue size, and concurrent players across global infrastructure.

Expanded Splunk from an unmaintained audit tool into a core component of company-wide monitoring. Deployed across all infrastructure and worked with Splunk professional services to build operational dashboards.

Administered and developed a homebrewed SSIS-based ETL framework running 600+ active jobs for insurance claims processing. Managed critical workflows for customer data acquisition, claims, and fulfillment. Built a knowledge management system and helped manage an offshore team in India.

A SaaS tool for leaders to visualize alignment between company goals and the work teams are actually doing, identify bottlenecks and risks, and manage workload capacity. Connects to project management tools (Jira, Linear, etc.) and surfaces where planned work drifts from strategic objectives — giving leaders real visibility without micromanaging. Designed with a bidirectional sync strategy to meet teams where they already work.

An AI-powered radiology training platform that helps medical students and radiologists sharpen their diagnostic skills. Students write diagnostic reports for chest X-rays, and the system compares them against expert ground truth using multi-step LLM analysis — delivering finding-by-finding scoring, educational feedback, and actionable improvement suggestions in real time via streaming. Built for the Spoon OS hackathon with CodeSeoul.

A web app for assisting medical imaging device technicians in China. Designed to run models like Qwen on local infrastructure or with remote models like Claude. Uses agent tools to lookup repair guidance documentation via RAG and web search.

Turn-based battle game where players would pit classic monsters against each other. Included account management, matchmaking, gameplay, and stats. Never released.