Tasnim Kabir

Multimodal QA: Audio + text reasoning datasets.
Earudite: Crowdsourced data collection platform for ASR training.
Geoguessr RL model: Reasoning model for geography game.

About Me

Tasnim Kabir is a Ph.D. researcher in Computer Science at the University of Maryland, working at the intersection of multimodal machine learning, question answering, and data-centric AI. Their research focuses on building robust benchmarks and datasets that enable models to reason beyond text, particularly in audio and real-world settings. Tasnim has contributed to research on transforming and scaling question answering data, including work on generating naturalistic QA datasets that improve model generalization . They have also developed systems and datasets such as AudioQA, targeting audio-grounded reasoning and evaluation for modern multimodal models. Their work emphasizes rigorous evaluation, failure analysis, and scalable data pipelines, with experience spanning dataset construction, model assessment, and end-to-end system development. In addition to research, Tasnim has led collaborative projects and mentored students, contributing to both academic and applied machine learning efforts. They are particularly interested in building reliable, real-world AI systems that integrate multiple modalities and move beyond purely text-based reasoning.