Built an Indonesian version of VQA-RAD and benchmarked general, multilingual and medical VLMs on English vs. Bahasa Indonesia prompts. Accuracy drops 8–25% in Indonesian, with systematic errors such as yes/no flips, left/right mix-ups and answers in the wrong language. Accepted at the CVPR 2026 MMFM-BIOMED Workshop. [paper]
A training-free subtitle translation pipeline that adds dialogue history and terminology to the prompt, generates diverse candidates, filters off-target outputs and reranks with quality estimation. Placed 3rd (Indonesian) and 5th (English). [paper]
Instruction-tuned the KomdigiUB-8B-Base Indonesian LLM for Indonesian legal compliance and digital-risk monitoring (+2% on IFEval), using synthetic instructions generated with Gemini and filtered by an LLM-as-a-judge. Built with the AI Talent Factory, Brawijaya University × Komdigi. [model]
Lab research funded by a DIPA FILKOM 2025 grant: 7 loss functions compared across 3 multigraph fraud-detection benchmarks.
Evaluated training-free detectors for AI-generated video, combining spatial-temporal likelihoods, second-order features, forensic augmentation and face-consistency matching.
A WhatsApp assistant backed by a medical Graph-RAG pipeline with query rewriting and context-based database routing.
Multi-head MobileNetV2 that recognizes crop diseases and nutrient deficiencies from phone photos (~90% accuracy), built at the Basic Computing Community.
Transcribes doctor–patient consultations with Whisper and summarizes them with ProphetNet to cut clinical paperwork. [paper]
Topic classification of tweets about the 2024 Indonesian election, handling class imbalance with a balanced-class approach and Optuna tuning. [code]