Every decision starts
as a single data point.
Follow this one.
Sunmin (Mina) Kim — M.S. Applied Data Science @ USC (Class of 2028), seeking a Summer 2027 internship. I turn messy data into decisions companies can act on.
3h → 0 daily manual search automated · 490K+ sale prices matched · 208/210 daily runs completed
It runs in production, every day.
A colleague at Global Vision spent 3 hours a day searching a government auction site by hand. I built the company's first data pipeline, alone: a daily ETL with a quality gate, an API, and a dashboard that ranks every listing against real sale prices. I also built the sales team's call management platform, web and Android. The pipeline still runs every day: open the live read-only demo ↗ (free hosting, first load can take about a minute).
Onbid Auction Pipeline
It started with one person spending 3 hours a day searching by hand. Now a daily 5-stage ETL does it overnight: a quality gate that halts bad runs, a web backfill that restored 17K bid rounds the API deletes, and tiered matching to 490K+ real-transaction prices. 208 of 210 daily runs completed.
GV TM — Call Management Platform
Outbound-call management system used daily by sales teams: admin web dashboard, agent mobile app, and async backend — built solo, end to end.
Before that, it became a metric.
Semo's AI diagnostic platform was live, but nobody could see it in numbers. I wrote the GA4 event spec from a blank page and checked the developers' implementation, then built BigQuery reporting on the raw export (retention, payment conversion) behind a Looker Studio dashboard the executives read every week. In beta I tested 4 hypotheses and turned the findings into prioritized fixes.
It learned new shapes.
Echo Rings: Coordinated Posting on a Graph
129K stock-board posts turned into a graph that links accounts posting the same sentences and group-chat invites. 553 accounts fall into one connected ring. The explorer runs the graph database in your browser, and an MCP server lets an AI agent query the evidence.
Document Image Classification
Team lead. 17-class document classification with an Optuna-tuned ensemble of ConvNeXt, EfficientNet, and Swin plus TTA — F1 0.97+, 2nd place overall.
Dialogue Summarization
Team lead of six. Summarized 12K everyday dialogues, scored on ROUGE: fine-tuned KoBART/T5 with LLM-based data augmentation and Optuna tuning.
CTR Prediction, Toss ML Next Challenge
Team lead of three. Ranked 70th of about 700 teams on 10M+ rows of ad data with a 1:52 class imbalance: a chunked Polars preprocessing pipeline and a DCN-v3 model, tracked in W&B.
MLOps Movie Recommendation
Team of four. I owned the data layer: TMDB API ingestion, PostgreSQL schema design, and S3 Parquet backups, inside an Airflow pipeline that automates training and inference on AWS.
More: Seoul Metro Deficit Analysis (K-Means station clustering, fare-policy proposal) · Commerce Marketing AI Agent (LangChain · FastAPI · SQLite FTS5 RAG)
Off the clock, it still ships.
Things I build for fun — live, real users, zero deadlines.
Idala — AI Nail Art Studio
Drop in photos you love, get three press-on nail sets, then see the one you pick on a photo of your own hand. A vision model inspects every set before you see it, and 419 automated tests guard the pipeline.
yzrecipe — Viral Recipe Tracker
Fully automated pipeline tracking Korea's trending recipes: scheduled collection every 3 hours, two-speed curation, zero-ops static publishing.
StockMind — Do Comments Move Stocks?
An hourly feed of 100K+ Yahoo Finance posts on 15 tickers with retries, backfill, and row-count checks, then 16 experiments with time-split evaluation and bootstrap confidence intervals. The honest result: posts add no signal beyond trading history.
myquant — Rule-Based Quant Agent
A deterministic stock-strategy research tool rebuilt from TradingAgents and ai-hedge-fund: hybrid single-ticker agent, walk-forward validation, cross-sectional long-short — stdlib only, LLM optional.
Sharpening the toolkit.
M.S. in Applied Data Science at USC (2026–2028) — data infrastructure, large-scale ML, and everything in between. 🏆 Minister of SMEs and Startups Award, Korea (2023) · 🎓 B.B.A. Global Business and Technology with a double major in AI Convergence, HUFS. Valedictorian (1 of 200+).