projects
Research and production work, most recent first. Numbers are measured on held-out data, not training splits.
anomaly-detection
2025–2026one-class intrusion detection for IoMT network traffic (thesis)
- Developed a CNN-DROCC one-class model for anomaly detection in Internet of Medical Things network traffic.
- Achieved 0.92 accuracy on CIC-IoMT-2024 and 0.97 on WUSTL-EHMS-2020.
- Converted tabular traffic features into structured 6×6 grayscale representations.
- Applied localized adversarial training to learn compact normal-data boundaries and prevent representation collapse.
- PyTorch
- CNN-DROCC
- Python
- CIC-IoMT-2024
- WUSTL-EHMS-2020
multimodal QR code phishing detection
- Engineered a three-branch neural network combining a Transformer lexical encoder with URL and QR structural features.
- Achieved 0.96 binary classification accuracy on an imbalanced phishing dataset.
- Designed cross-modal attention across 32-dimensional branch embeddings to learn interactions between modalities.
- PyTorch
- Transformer
- MLP
- Cross-Modal Attention
multi-user document RAG platform
- Built a multi-user document RAG application with secure PDF, DOCX, and URL ingestion; hybrid retrieval, reranking, and source-cited Gemini responses.
- Implemented JWT authentication, capsule-level data isolation, background ingestion workers, persistent chat history, automatic capsule summaries, and continuation handling for long LLM outputs.
- Deployed on Google Cloud with PostgreSQL, Redis, and Qdrant-backed services configured for production use.
- FastAPI
- PostgreSQL
- Redis
- Qdrant
- Gemini
- Google Cloud