High school student in Morocco · AI tooling & applied machine learning
Applied ML · AI systems · Reproducible experiments
I build practical AI tools and explore how models generalize to unfamiliar data. My work brings together software engineering, self-hosted inference, and controlled experiments. I’m looking to deepen my research skills through mentorship and collaboration.
Research · LinkScribe · ORCID
Question: does training on more regions change the advantage of multispectral imagery over RGB, independently of total training-data volume?
| Configs | Models | Evaluations |
Figure from the study repository: performance as geographic training diversity changes. Open the image for the full-resolution figure.
What the evidence supports
Multispectral models outperformed RGB at every tested region count. The 95% bootstrap interval for the interaction slope included zero, so the study found no clear evidence that this advantage changed systematically with geographic training diversity.
Research status: manuscript in preparation; not yet published.
Experimental design & reproducibility
- Fixed total budget: 48 training patches per model, redistributed across regions.
- 1, 2, 4, or 8 training regions, enumerating all 107 subsets of the eight-region roster.
- RGB and multispectral inputs use paired, deterministic patch sampling.
- Two modalities and three seeds: 107 × 2 × 3 = 642 models.
- Two validation regions for model selection; four separate regions for final testing.
- Frozen configurations, normalization artifacts, checkpoint hashes, and final results are documented in the repository.
|
AI tooling / local inference English transcripts from public video links, served to AI agents through an authenticated FastAPI queue. Speech recognition runs locally with whisper.cpp on an ARM VPS. Python · FastAPI · SQLite · FFmpeg |
Device control / distributed systems Google Home controls for Windows. A .NET agent connects outward to Cloudflare and executes a defined set of actions without exposing a remote terminal. C# · .NET · Cloudflare · TypeScript |
Also built: RemoteCraft, a Minecraft server dashboard, and Lumen, a native Rust music player for Windows and Linux.
How to design stronger model evaluations, distinguish data effects from model improvements, and make experiments easier to reproduce. I’m interested in research mentorship where I can build on both the software and experimental sides of my work.
Tools used: Python, PyTorch, rasterio, FastAPI, SQLite, Rust, C#, Linux, GitHub Actions.