Dark Phoenix Next.js, Modal, WhisperX, Gemini, FFmpeg
- Situation: Long videos are slow to repurpose into short vertical clips.
- Task: Build a pipeline that automates the conversion.
- Action: I built it in five days and fixed three distributed-systems failures: a silent event queue, a credential misconfiguration, and a race between the GPU result and the database write.
- Result: A working end-to-end pipeline from long video to short clips.
TWIN: Multimodal Deepfake Detection PyTorch, T5, Swin Transformer, LIME
- Situation: Detecting manipulated media needs both text and image evidence.
- Task: Build a detector that uses both modalities and explains its output.
- Action: I fused a T5 text model and a Swin Transformer image model with cross-attention, and added LIME explanations.
- Result: The model reached 93.7% accuracy, 14.2% above baselines. I presented it at Sigma Xi and placed 3rd Runner-Up.
Open Source:
- Cal.diy: Submitted a pull request adding conditional booking questions.
- OpenRootDev: Joined the community and contribute to Lysis, a beginner-friendly issue ranker. Lysis ranks a repository's open issues by how likely a first-time contributor is to get a PR merged. Because
good first issuelabels are often stale, it learns from issues that first-time contributors actually closed.
Contact: hvora@umassd.edu

