Papers and projects.

A web interface for browsing the Masader catalogue of Arabic NLP datasets — exploration, filtering, and an API over the backend.

Most trackers model each target in isolation, which breaks down when objects look alike. We introduce a multi-track pooling module that updates appearance memory across all tracks at once, with a matching training strategy that generates hard tracking episodes online — real-time state-of-the-art on public MOT benchmarks under a bilinear-LSTM framework.

GPU expansion cards hit real limits once they have to move large data across the system bus. We evaluated several GPU architectures on massive audio, video, and DNA sequence workloads at Oregon State's CGRB, and showed where CPU–GPU integration is the actual bottleneck.

On the Passive Violation of Expectation team under Dr. Alan Fern. Built a tracker on Bilinear-LSTM and a Siamese network, a role-assignment module to decide which objects matter, and a rule-based reasoning agent that judges whether a scene is plausible.

Automatically updates missing module versions in requirements.txt to the latest.

A Discord and Twitch bot running style transfer, English and Arabic text completion, Arabic poetry generation, English QA, and TTS — built for Twitch streamers.

Took the model from 23.01% to 41.57% by swapping the image embedding for pre-trained CLIP and the word embedding for GloVe. Built a web application on top of it.

An ML pipeline that auto-labelled over 100,000 bird sound samples using hidden Markov models, FFT, k-means and bootstrapping, then a deep learning classifier to detect illness from the calls — deployed to an NVIDIA Jetson Nano with TensorRT.

An image labelling tool that took throughput from 300 to 3,000 images per week. Used across multiple labs, including fieldwork in Africa with no internet access.

A social network analysis system for understanding how players interact during a match.

A computer vision system for the City of Portland that detects and obfuscates pedestrians' faces in real time, then tracks movement to study pedestrian and traffic behaviour for road safety.

A site that races different hardware against each other running the lab's deep learning models. Sponsored by Tech Data, IBM, NVIDIA and OpenPower, and featured at IBM Think 2019.

Deep learning models that classify owl species from audio.

Deep learning models for detecting plankton in imagery data.

A CNN–RNN model that generates captions from images: a convolutional network for feature extraction, an LSTM for generating the caption.

A web app streaming live video from a Raspberry Pi-controlled RC car, plus the intelligent agent structure and model behind it.

A card game built with a team. Won Best Web Application for Software Engineering I.

An early version of this site, plus an NGINX config that redirects HTTP to HTTPS and maps networks.

Hunt the Wumpus with a dynamic grid size, and an embedded agent that solves it using probability distribution methods.