I'm a master's student in AI and Machine Learning at PAUSTI in Nairobi, and until January 2027 a special research student in the Dependable Systems Laboratory at Hiroshima University, with Assoc. Prof. Junjun Zheng. Before that I studied Statistics and Programming at Kenyatta University and worked as a data analyst and cybersecurity associate in Nairobi.
I work where machine learning, security and low-level systems meet. My thesis is malware detection that reads Windows programs byte by byte and can tell when it is looking at a family it has never seen. Further down the stack, I'm interested in the hardware models run on: side-channel and fault attacks on GPUs, Rowhammer in GPU memory, and keeping a model's weights and inputs from leaking.
News
- 17 oct 2026Talk on open-world malware detection at the Chugoku-section joint conference of electrical and information engineers, Hiroshima (programme).
- sep 2026Wrote about how GPU memory works and Rowhammer.
- aug 2026Started a research stay in the Dependable Systems Laboratory, Hiroshima University.
- jun 2025Started the master's at PAUSTI on a Pan African University scholarship.
Research
Open-world malware detection master's thesis
A malware classifier gives every new family the name of an old one. I'm building detection that can also say "unknown", then learn the new family from a few examples. An encoder reads a Windows program region by region, as the PE format lays it out, and is pretrained without labels on the 9.9 million binaries of SOREL-20M. Each known family is described by a few prototypes, and a file far from all of them is rejected.
With Dr. Albert Njoroge Kahira and Dr. Mulang' Onando (PAUSTI) and Assoc. Prof. Junjun Zheng (Hiroshima). Talk: Open-World Windows Malware Detection with PE-Aware Hierarchical Embeddings and Few-Shot Adaptation, R. A. Zome, J. Zheng, H. Okamura, T. Dohi, Hiroshima, October 2026.
Federated intrusion detection for vehicles in progress
GraMa reads in-vehicle CAN traffic as a graph of message IDs with graph attention, follows it over time with a Mamba model, and trains across vehicles without sharing their data. The server clusters the vehicles' updates with HDBSCAN to keep poisoned ones out. code
Projects
- raw-pe, a Win64 executable written byte by byte in NASM, no linker. codewrite-up
- sco-pe, a PE parser in C with bounds checks for malformed files. code
- Explainable deepfake detection, an Xception classifier that explains each verdict with retrieval. democode
- Phish-Transformer, a 45K-parameter URL classifier behind a Chrome extension. codemodel
- ML on Kubernetes, a FastAPI model on GKE with CI/CD and monitoring. code
- Swahili GPT, a character-level GPT trained from scratch on Swahili news. model
Writing
Scholarships
- 2025–2027Pan African University scholarship, African Union (master's)
- 2019–2023County government scholarship, Kenya, for high school and again for the B.Sc. at Kenyatta University