CV

Virginia Tech

Jan 2023 – Dec 2027 (expected)

Ph.D. Candidate, Computer Science. Advisor: Dr. Chris Thomas. GPA 4.0/4.0. Amazon Fellow.

Jordan University of Science and Technology

Feb 2020 – Jun 2022

M.S., Data Science. Advisor: Prof. Rehab Duwairi. GPA 4.26/4.30. Full scholarship (Outstanding Graduate Student Award).

Jordan University of Science and Technology

Feb 2016 – Jan 2020

B.S., Computer Science. Advisor: Dr. Malak Abdullah. GPA 4.04/4.20. Ranked first of 62.

Graduate Research Assistant, Virginia Tech

Mar 2023 – Present

• Lenses (CVPR 2026). Built a polysemous image–text benchmark (105K images; 732K automatically generated, model-verified captions across five interpretive lenses) and a lens-token multi-embedding model that outperforms the evaluated baselines on literal and non-literal retrieval (figurative image-to-text R@1: 56.0 vs. 28.2 for the best evaluated zero-shot baseline). Led problem formulation, dataset design, modeling, experiments, and writing.
• MaxMatch (ACL 2025). Maximal Pair Assignment Similarity with two diversity losses that prevent set collapse; state of the art on Flickr30K and COCO without external data (+27.3 RSUM on Flickr30K and +21.3 on COCO 5K over SetDiv). Used as a baseline by independent researchers. Led method design, implementation, and evaluation.
• Length bias in vision–language models (EMNLP 2026, SAC Highlights). Showed that models prefer longer answer options (accuracy drops by 2–38 percentage points when the correct answer is shorter), traced the bias to late-layer suppression of visual evidence, and showed that inference-time fixes alone are insufficient. Contributed to study design, experiments, and analysis.
• Mind the Channel (under review). Test-time inference for character-corrupted queries with frozen models. On the COCO-C and Flickr-C synthetic-corruption benchmarks, recovers 92% and 94% of the retrieval accuracy lost to corruption.
• Video and event understanding. NEST (EMNLP 2026; 1,005 full-length movies with about 103 grounded narrative events each), M³EC (EMNLP 2026; events across news text, video, and audio), ENTER (MAR Workshop @ NeurIPS 2024, Spotlight, with Columbia), and TimeCheck (under review). Contributed to task design, experiments, and analysis.
• EchoBelief (under review), with Dr. Alaa Algargoosh. A robot maps rooms from echoes alone, deciding where to move, when to listen, and when to stop. In 355 simulated rooms with known poses, it reaches 90% wall recall in 70.1% of episodes vs. 49.4–52.0% for three tuned planners, using about half the sensing. Built EchoRooms (4,026 rooms, 1.58M room impulse responses). Led the project.
• Part-level visual grounding (ECCV 2026). Coarse-to-fine object-to-part grounding with reinforcement learning; supported by the NSF HDR Imageomics Institute. Contributed to method design, experiments, and analysis.
• Surgical instrument recognition (CVPR 2025 FGVC Workshop, with Carilion Clinic). Real-time classifier for 95 ultra-fine-grained instrument types; a prototype evaluated in Carilion's sterile processing department, where six technicians rated it 81.7 on the System Usability Scale. Contributed to modeling, experiments, and the user study.
• Envision (ongoing). A world-model approach to composed image retrieval that imagines the target image before searching for it. Leading the project.

Research Assistant and Undergraduate Researcher, JUST

2019 – Nov 2022

• Subjective and adversarial language. Emotion detection in dialogue (EmoDet, EmoDet2; an independent group reused the EmoDet2 design in six follow-up studies), propaganda detection in news (NLP4IF 2019, ranked 7th of 25 teams in sentence-level detection), and humor and offense rating (SemEval 2021, 0.954 F1 on humor detection).
• Machine-learning security. LightGBM pipeline for multi-family Android malware classification on CICMalDroid 2020 (Procedia Computer Science 2023).
• So2al-wa-Gwab. A 10K-pair natively written Arabic QA dataset; led a three-person annotation team with three-stage quality control (ACM TALLIP 2023).

Machine Learning Engineer Intern, Adobe

May – Aug 2026

San Jose, CA

• Built a large-scale multi-persona image annotation pipeline for image-quality evaluation (filtering, tagging, synthetic annotation, quality validation).
• Ran a visual-embedding similarity study and developed a DINOv3 + Transformer quantile-regression model that learns quality distributions of visual attributes from Adobe Stock images.
• Introduced distortion-aware training to predict high-quality targets for degraded images, improving Q-Align on MIT-Adobe FiveK and outperforming the evaluated baselines.