Projects

In this work for the MBZUAI Trustworthy NLP course project, I aim to mechanistically understand the factual hallucination in a multilingual setting.  In particular, I examined two types of hallucination: knowledge enrichment and answer extraction hallucination

In the MBZUAI Advanced NLP course, we analyze how SLMs distinguish between benign, harmful, and jailbreak inputs. By identifying and manipulating key directions in the model’s internal representations, we can flip its behavior with a single targeted intervention.

Our project examines how Whisper-based ASR models handle noisy real-world conditions in low-resource Indonesian languages. We fine-tuned models using SpecAugment and noise-aware training, showing major gains in recognition accuracy under low signal-to-noise ratios. Our results reveal language-specific error patterns and demonstrate the importance of noise-robust and dialect-aware adaptation for regional ASR systems

Implementing CNN-RNN-based lane detection for undergraduate final project