I turn vague problems into testable questions.
My work spans full-stack web development, machine learning, and open source. When working with AI, I make the scope and constraints explicit, distinguish facts from assumptions, and verify proposed solutions through tests, logs, and measurements. I share what I learn and refine my decisions through experiments and code review, weighing performance, complexity, and reliability.
π« Open source Β· Fedify
Fedify is a TypeScript framework for building federated services with ActivityPub.
Latest pull request Β· Awaiting review
Added Ed25519 request signing for RFC 9421, including retries for Accept-Signature challenges. A round-trip test signs a request and checks that Fedify can verify it. #1438
Merged contributions
| Area | Contribution | Pull requests |
|---|---|---|
| WebFinger diagnostics | Exposed opt-in lookup diagnostics through withDiagnostics and added distinct CLI messages for missing resources, invalid responses, and network or server errors. |
#1296 |
| Debugger dashboard | Added trace and log filters to the dashboard and its JSON APIs, and reworked polling to detect changes in the filtered trace list. | #1204 |
| Static analysis | Rewrote the outbox delivery rule using path-aware AST analysis and added a separate rule for unawaited delivery promises. | #1050, #1067 |
| Actor validation | Added a preferredUsername lint rule and corrected shared handling of Tombstone returns and preferredUsernames initializers. |
#1022 |
| Regression tests | Added tests for the Express and h3 integrations and tested KV expiry with Miniflare bindings. | #985, #1007, #1021 |
For actor validation, invalid WebFinger responses, and Ed25519 signing, I confirmed that targeted regression tests failed on the original code and passed with the fixes applied.
π Machine learning projects
β½ K League Pass End-Point Prediction
K League Γ University of Seoul Open AI Competition Β· Gold Award (2nd place)
Predicted the final pass destination from match-event sequences using a BiLSTM, spatial heatmaps, and dense supervision. The repository documents match-grouped validation, consistent decoding during training and inference, and reproducible fold artifacts.
π΄ Restaurant Menu Demand Forecasting
LG Aimers 7th Β· Team project Β· 46th on the private leaderboard
Forecasted seven days of menu sales from a 28-day history using an ensemble of tree models, an LSTM, and a seasonal baseline. Refactored the notebook into a CLI package, compared outputs on synthetic data, and documented the limits of the offline validation.
π§ Web development Β· VIBR
A social music service built with a four-person team. I worked across the frontend and backend on authentication, search, playback, feed consistency, and event logging, then revisited performance and reliability through reproducible measurements.
π± Tools I work with
Web
ML



