01 — Garuda Hacks 6.0 · 30-hour buildBackend Developer
Vika AI
Real-time multi-modal AI tutor and voice agent
An adaptive learning platform for students with ADHD that teaches to the VARK learning styles — visual, auditory, read/write — instead of forcing one mode on everyone.
- Built the Node.js, Express and MongoDB backend, and the API layer joining frontend, AI models and database.
- Orchestrated real-time sessions that sync Gemini 2.5 Flash verbal responses with live canvas visualisation commands.
- Shipped a live conversation engine on Google Cloud STT/TTS with custom voice-agent logic and an automated ADHD assessment flow.
TypeScriptNode.jsExpressSocket.IOMongoDBGemini 2.5 FlashDocker
02 — Machine Learning final projectData exploration, feature engineering, model tuning
SNBP Admission Prediction
Classification under severe class imbalance
Predicting outcomes in Indonesia's achievement-based university admission track. The real problem was not the model — it was a 90.9% class imbalance that made naive accuracy meaningless.
- Engineered an indeks prestasi feature combining academic and non-academic achievement into one signal.
- Demonstrated SMOTE oversampling materially outperformed random undersampling at recovering the minority class.
- Compared LightGBM, XGBoost and Random Forest, then tuned the leader with Optuna.
Best configuration — LightGBM + logistic regression, soft voting — reached 0.735 accuracy at 0.576 ROC AUC. Honest numbers: the imbalance limits how far this data goes.
PythonLightGBMXGBoostSMOTEOptunascikit-learn
03 — Research tooling · recongraph, reconfsm, reconformalDevOps Engineer & Core Python Developer
Forensic Timeline
DFIR tooling, from notebooks to distribution
Three digital-forensics research tools existed as loose scripts and notebooks. I turned them into installable, documented, continuously-tested Python packages.
- Refactored flat scripts and notebooks into modular packages managed through pyproject.toml.
- Built GitHub Actions workflows for matrix CI and automated release to PyPI on tag.
- Automated documentation builds via Read the Docs, and secured publishing with PyPI Trusted Publishers (OIDC) instead of long-lived tokens.
PythonGitHub ActionsPyPI OIDCRead the Docs