Ki-Ung Song
AI grounded in mathematics
Edge AI, AI for Science, AI for Finance
Seoul, South Korea
Growth Hackers in SNU
Business Data Professionals of SNU
Overview
- At Growth Hackers in SNU, I worked on industry-facing data projects with partner companies, translating messy business questions into workable modeling and evaluation problems.
- That work taught me early to see applied ML not as isolated model-building but as framing the right question, understanding the data, and deciding what result would actually be useful.
- As the 7th education team leader, I also redesigned the internal curriculum so newer members could learn data analysis and modeling in a more practical, structured way.
Selected Projects
RecSys Modeling with Educast
Oct 2020 - Dec 2020
-
Compared modeling directions for education-content recommendation:
- Deep generative recommenders I implemented in PyTorch
- Content-based methods built on embeddings and clustering
-
The focus shifted toward evaluation: which metric reflects a good recommendation in a given scenario?
- I led that analysis and argued that evaluation should depend on the recommendation context, not be one-size-fits-all.
- The main value was clarifying how to judge modeling and evaluation choices in applied recommendation.
Automating Delivery Area Assignment and Cargo Demand Prediction with Timfresh
Jul 2020 - Aug 2020
-
Delivery area assignment
- Manually assigned areas left similarly paid drivers with very different loads.
- As technical co-PM, I redrew areas by actual route distance from a map API and assigned them with clustering that also considered hierarchical structure.
- Comparing variance against the manual split showed the data could support more balanced, automated planning, though the system was not deployed.
-
Cargo demand forecasting
- I compared the DL models the company hoped for with ARIMA and SARIMA. Simple seasonal models still won.
- With few features and little history, I recommended fixing the data before adding model complexity.
- It taught me to tell whether the bottleneck is the model or the data.
RecSys Modeling with Mathpresso
Apr 2020 - Jun 2020
- Basic EDA surfaced the main issue before modeling did: after filtering, too little usable data remained for a strong recommender.
-
I still built a predictive recommender in TensorFlow and analyzed its embeddings.
- They were heavily biased toward a narrow set of data features.
-
The key outcome was not a better model but the conclusion that the data had to improve before more sophisticated ML would help.
- Communicating that back to the company was one of the most valuable parts of the project.
What It Taught Me
The lasting lesson was that applied ML starts with the data. Before model choice matters, you have to know what signal the data contains, what filtering removes, and whether what remains can support a meaningful decision. Better modeling only matters when the data can support it.