entry // front page
researcher, note-taker, occasional poem-hoarder
[research notes on vision, language, and cognition]
I’m an M.Phil. (Research) student in Computer Science & Engineering at HKUST, working across computer vision, HCI, and visualization. My work explores embodied cognition in expert chess, alongside research in Arabic NLP and resource-efficient speech models. I got my BSc in Computer Science with a Machine Learning concentration at Carnegie Mellon University, in Qatar/US, where I also taught and did research in CV and NLP.
[training, methods, and recurring obsessions]
My research bridges computer vision and HCI, studying embodied cognition in high-skill settings like expert chess. I build vision pipelines using pose estimation and identity tracking across temporal sequences, paired with behavioral surveys and gaze tracking. For analysis, I design interactive visualizations and temporal charts to surface patterns in movement, decision difficulty, and cognitive load.
I am also active in visualization research, specifically chart literacy studies, cross-cultural visualization design, and translation of visual encodings across locales. Previously, I worked on compute-efficient pretraining for speech language models using compression-based representations and Arabic NLP in low-resource settings with cross-prompt evaluation.
Computer Vision Based Chess Analytics
I use vision and visualization to study embodied cognition and decision-making in expert play.
DataVisards Lab • Advisor: Dr. Arpit Narechania
Optimizing Speech Language Models
Focused on compute-aware pretraining and efficient representation learning.
ML & Signal Processing Lab • Advisor: Dr. Bhiksha Raj
entry // ongoing
I lift lids. I want to know what machinery hums underneath. In systems, in people, in myself.
I don’t fit a label. I like structure but resist boxes. Some days I chase multimodal models. Other days, prehistoric periods or plants I can name by sight. I move between them.
I write to make sense of things: research notes, sometimes poems. Not everything that sounds right means something. I’ve made peace with that.
I love teaching. That quiet moment when something clicks for someone else. I could chase that for the rest of my life.
Part of me wants a hut and solitude. The other wants to build, study, and stay in conversation. This website is the compromise.
[ongoing curiosities, published detours, working theories]
a few of the things I’ve built while trying to understand something properly
[in progress]
I combine livestream video, Stockfish-derived difficulty, and timing signals to study expert decision-making in chess. I built a behavior pipeline with 3D pose and mesh estimation, plus identity tracking. Then I model movement signatures over time.
[under revision]
I built a cross-prompt evaluation setup over QAES and CAST. I implemented PAIR-Net with prompt-conditioned cross-attention and caching, and ALiF-Net for Arabic linguistic features. Then I ran augmentation experiments to push agreement closer to human scoring.
[published]
Thermal images are noisy and unfriendly. I tested classic keypoint detectors and descriptors for thermal registration under domain shift, including SIFT, SURF, ORB, BRISK, and AKAZE. The result is a practical benchmark with clear winners, published at IEEE VCIP 2024.
[thesis work]
I explored resource-efficient pretraining strategies for speech language models. I prototyped compression-based representations using deep compression autoencoders. I also built synthetic supervision to model speech errors at scale.
[prototype]
I led a team building a multilingual, real-time captioning prototype using Azure and PowerApps. We added accessibility features like text-to-speech and translation. It earned 2nd place in a hackathon, which was a fun week.
[engineering detour]
I built a scalable backend using .NET 8 and MongoDB. The design uses modular services and DTOs, plus unit tests to keep it sane. I also Dockerized deployment and tuned it for high-concurrency order and inventory workflows.
[margins, drafts, fragments]
I write research notes when I’m trying to understand something. I write poems when I’m trying to say it cleanly.
[where the work happened]
I work on embodied cognition in expert chess. I built a CV pipeline with 3D pose/mesh estimation and identity tracking. I model behavioral dynamics over time and connect movement patterns to decision difficulty. A CHI manuscript is in progress.
I work on cross-lingual Arabic automated essay scoring. This includes benchmark construction, model implementation (PAIR-Net and ALiF-Net), and experiments aimed at improving agreement with human raters.
I wrote my senior honors thesis on compute-aware pretraining for speech language models. I explored compression-based representations and built scalable synthetic supervision for speech error modeling.
I studied descriptor robustness in thermal imagery for solar panel stitching. I implemented and benchmarked classic methods, then wrote and presented the work as first author at IEEE VCIP 2024.
I led recitations and supported implementation-heavy assignments for 250+ students. I taught core ideas behind Transformers and LLMs, CNNs, and autoencoders. I also answered a lot of “why is my loss NaN” questions.
I built backend services with .NET 8 and MongoDB, using a modular architecture and unit tests. I shipped Dockerized deployments and was recognized for “Most Promising Project.”
I led a team building a multilingual real-time captioning prototype using Azure and PowerApps. We placed 2nd in a hackathon.
I used STIX and knowledge-graph methods to structure cyber threat intelligence. I also curated and benchmarked CTI datasets.
I built segmentation and detection pipelines for surgical instrument detection using SAM, SegFormer, and YOLOv8. I tuned preprocessing and improved performance in real operating room data.
[open line]
letters, questions, and curious detours are welcome
if you’d like to send a note, you can begin with:
I’m usually happy to hear about research, writing, collaboration, odd little questions, or anything that feels like it belongs in the margins.
suggested sign-off, entirely optional:
with curiosity,