Computer Science @ Universitas Gadjah Mada · causal RL researcher · I build quiet, local-first desktop apps
I split my time between two things:
- Research on causal machine learning and reinforcement learning: learning decision policies from observational data without fooling yourself about confounders. Most recently as a research intern at Constructor University (Bremen, Germany), working on low-emission driving policies for urban logistics.
- Software that I want to use every day: fast, keyboard-friendly desktop apps that keep your data on your machine. Mostly TypeScript + Svelte 5 on Electrobun, with Rust and Swift where it counts.
Dictation, file transcription and meeting notes for macOS. Speech runs on-device (Parakeet, Whisper, and more on Metal); cloud and AI providers are opt-in. Hold fn, talk, release.
|
A modern, quiet IRC client for macOS, Windows and Linux. Full IRCv3 (history, replies, reactions, read markers), first-class soju bouncer support, and searchable local history.
|
A Plex client that plays everything in mpv, with your own config, shaders and scripts. Ships as a desktop app and a keyboard-driven Rust TUI with Miller columns and a scriptable
|
A desktop listening room for Audiobookshelf: a chapter-aware gapless player, mini player, EPUB/PDF reader, podcast management, offline downloads, and progress that survives crashes without ever overwriting newer progress from your phone.
A federated blogging platform: independent instances share posts as peers, with
|
| Project | What it is |
|---|---|
| Causal-RL | Confounder-Aware Eco-Driving Decision Support for Simulated Urban Logistics (first author). Confounder-aware Fitted Q-Iteration on a synthetic benchmark of 96,000 decision epochs, with leakage checks, support-constrained recommendations and doubly robust off-policy evaluation. |
| Neural medical search | Sparse retrieval with Pyserini + RM3 query expansion, re-ranked with ClinicalBERT. |
| resume-bert-project | Resume classification across 24 occupations: hybrid TF-IDF/SVD + transformer, a TrOCR path for scanned resumes, served with FastAPI on Hugging Face Spaces. |
| tuberculosis | TB detection from chest X-rays: classical lung segmentation + feature extraction + Random Forest (~90% accuracy), with a React front end. |
Interests: causal ML · reinforcement learning · intelligent transportation · sustainable logistics · robotics · local-first software
Usually shipping more than one desktop app at once. Say hi if any of them are useful to you.