$ portfolio
A curated list of my technical projects and research experience.
$ projects --selected
Some of my most recent projects.
Wafflr
Wafflr is a private weekly video ritual for close friends who want to stay in touch without the noise of group chats or performative social media. Friends share quick 2-minute updates in small groups, creating a low-pressure way to feel present in each other's everyday lives.
dex.ai
Video-to-3D reconstruction platform using photogrammetry, NeRFs, and Gaussian Splatting with on-demand GPU processing.
PennOS
Unix-like operating system with a shell, scheduler, process lifecycle management, and file system semantics.
RISC-V Pipelined Processor
Five-stage RV32IM processor with bypassing, stalls, branch prediction, and an AXI4-Lite memory interface.
$ writing --research
Accuracy and Fairness of Facial Recognition Technology in Low-Quality Police Images: An Experiment With Synthetic Faces
Studied how contrast, brightness, motion blur, pose, and resolution affect facial recognition in police-style images. The experiment used StyleGAN3 synthetic faces, FairFace demographic labels, and ArcFace-based identification to measure accuracy and demographic disparities under controlled degradation.
MoE-Beyond: Learning-Based Expert Activation Prediction on Edge Devices
Built a lightweight transformer predictor for which experts a sparse MoE model will activate during autoregressive decoding. Trained on expert activation traces from DeepSeek-V2-Lite, the system targets edge-device serving by improving GPU expert-cache hit rates under tight memory constraints.
Definitions, formulas, and simulated examples for plagiarism detection with FAIR metrics
Extended the FAIR family of citation and attribution metrics with ratio-based formulas for detecting suspicious attribution patterns in scholarly writing. The work uses simulated examples to reason about plagiarism screening before validation with human review and automated semantic analysis.