Ryan Leem

I wanted to make it easier to find existing GitHub issues when I only knew how to describe a bug in plain language. I built a semantic search tool that indexes 11,336 bug reports and can retrieve current public issues through the GitHub API.

I collected and cleaned the issue data, generated sentence embeddings, stored them in PostgreSQL with pgvector, and added title-aware reranking with a FastAPI search path. On 20 paraphrased tests, the top result improved from 70% to 80%, and I learned that titles carry useful signal when an issue body is vague.

I wanted to find which bank account applications looked most likely to be fraudulent instead of treating every application the same. I analyzed 1 million applications in PostgreSQL and built a rule-based risk score that groups accounts by warning signs.

I wrote the SQL with CTEs, indexed the fields used by the score, and checked query plans with EXPLAIN ANALYZE. The highest-risk segment had an 11.74% fraud rate—more than 10x the baseline—and the indexed query ran in about 6–7 ms. This showed me how much a simple, explainable score can reveal before adding a more complex model.

I wanted to turn a slow, repetitive monthly banking report into a workflow that could handle more rows than Excel can hold in one worksheet. I built a reporting process that aggregates multi-million-row transaction data before loading the results into an Excel dashboard.

I wrote the Power Query transformations, set up the PivotTables and PivotCharts, added VBA for one-click refreshes, and used a CSV mapping layer to keep the inputs reusable. The result made the report repeatable without forcing the full raw dataset into Excel, and it taught me to solve scale limits before they become spreadsheet problems.

I wanted to understand where customers were dropping out of an e-commerce funnel and whether brands with more traffic were also converting better. I analyzed more than 500,000 interactions in R and compared funnel steps and conversion rates across brands.

I cleaned and grouped the event data with tidyverse, built funnel charts with ggplot2, and ran a regression to test the relationship between traffic and conversion before filtering to high-traffic brands. The regression had an R² of 0.0845 and a p-value of 0.335, so traffic alone did not explain conversion well; comparing brands on more than volume gave me a more careful way to read the funnel.