I'm a Computer Science student at UW–Madison (Fall 2026).
I focus on Applied AI and backend-driven full-stack development, building production systems that combine LLMs, retrieval pipelines, and scalable cloud infrastructure.
My work spans RAG systems, multimodal AI pipelines, and distributed backend architectures.
raphael.oh0316@gmail.com · Portfolio · LinkedIn
Languages: Python, Java, TypeScript, C++
Frameworks: FastAPI, Spring Boot, Node.js (Express), Next.js, React Native (Expo), Django
Databases: PostgreSQL, MySQL, MongoDB, VectorDB(Pinecone), Neo4j, Redis
Infra/Data: Docker, Nginx, AWS (EC2/S3), GCP Dataproc, Google Cloud Run, Kafka, Celery, Firebase Authentication & Realtime DB, Apache Spark
AI/ML: RAG (Hybrid Retrieval, BM25, Pinecone, RRF), LLM Orchestration, Model Evaluation, LangChain, LangGraph, PyTorch, scikit-lear, Hugging Face (Transformers), Models(CLIP/CLAP/EmoCLIP/ImageBind, YOLOv8, MoveNet)
MLOps: Vector DB pipelines, ETL pipelines, batch/stream data ingestion, distributed inference (Celery + Redis), model serving, monitoring (Prometheus + Grafana)
A real-time AI coach analyzing biomechanics from video
- Fine-tuned YOLOv8 + MoveNet Thunder, achieving 94% detection accuracy and ~60ms inference
- Built 6-phase shot segmentation (setup → loading → rising → release → follow-through)
→ 91% segmentation accuracy across 100+ labeled videos - Engineered DTW-based similarity scoring with kinematic features (angles, timing, velocities)
- Built a FastAPI backend + React Native app, with caching & deduplication for real-time mobile use
- Integrated an LLM feedback module that generates personalized, coach-style tips
- Tech: Python, TensorFlow Hub, PyTorch, OpenCV, FastAPI, Docker, React Native
IOS Appstore · GitHub · Portfolio
An LLM-powered system that extracts, normalizes, and manages story settings
- Built an LLM-driven extractor for characters, places, items, relationships, skills
- Designed self-referential PostgreSQL schema for hierarchical world knowledge
- Detects setting conflicts using embedding similarity + coreference resolution
- Implements temporal attribute scoping for multi-chapter consistency checks
- Built with FastAPI + PostgreSQL + Alembic + Docker
A structured tabular ML system for movie revenue forecasting
- Engineered 58+ features from nested JSON metadata (cast, genres, release timing), expanding 23 raw columns → 81 structured features with log-scaling and categorical preprocessing
- Designed and compared Linear Regression, XGBoost, and a custom Residual MLP (BatchNorm + Dropout)
- Conducted ablation studies across 10 configurations to analyze regularization and bias–variance trade-offs
- Improved RMSLE from 2.10 (baseline) to 0.18, with tuned XGBoost outperforming deep learning models on structured tabular data
- Built with Python + PyTorch + XGBoost + Scikit-learn
A social–academic app for 500+ students (Next.js + React Native)
- Integrated Madgrades API → PostgreSQL ingestion (900k+ records)
→ parallel batching, idempotent upserts → 33% faster pipeline - Auto-generated semester-based course chatrooms using Firebase Realtime DB
- Verified-review system: students staying in chat all semester get a verified badge
- Built mobile client with React Native; backend on Next.js (App Router)
Mobile App · Backend · Portfolio
A dual-backend architecture for designers & clients
- Spring Boot (auth, users, projects) + Node.js (Socket.IO, Mediasoup)
- Supported 50+ concurrent sessions <200ms latency
- Deployed via Docker + Nginx reverse proxy with JWT and Redis token rotation
- Real-time whiteboard, multi-user video calls, shared workspace
Auth Server · Stream Server · Portfolio
Django-based lightweight reading app
- Upload & parse novel files with bookmarking, theme toggling, reading modes
- Built for mobile-first readers
RFID-based food identification assistant
- Used MongoDB + Node.js for food metadata & allergy tracking
- Built admin APIs, meal mapping, and RFID integration
- Android app built in Kotlin with accessible UI design
C++ educational game mixing word creation & RPG-like stat progression
- Led a 3-person team (PM role), scoped project, designed UI & mechanic flow
- Built alphabet → stat mapping system to deepen learning through gameplay
Award-winning assistive typing layout
- Designed directional Hangul layout for one-handed use
- User-tested iteration cycles; formal competition project
- Silver Prize, 2018 Seoul Creative IT Competition
- Multimodal AI (vision–language–audio)
- Retrieval systems (BM25, Pinecone, hybrid RAG)
- MLOps & large-scale inference pipelines
- Real-time collaboration software
- Accessibility & education-focused tech
Thanks for visiting!
If you're working on AI infra, multimodal systems, or scalable backend tools — I'd love to connect!