PhD researcher, Tianjin University
The model is rarely the hard part. The data is. I work on getting noisy biomedical signals to agree with each other, then checking whether they still agree on data the model has never seen.
I'm a PhD candidate in Information and Communication Engineering at Tianjin University, supervised by Prof. Zhang Tao. My research is on catching Parkinson's disease early, using signals that start shifting years before a clinical diagnosis is possible.
The problem with any single biomarker is that it is easy to fool. A model can score brilliantly on one kind of signal and still miss the disease completely, or latch onto some artifact that has nothing to do with the patient. So rather than trusting one signal on its own, I look for agreement across several before believing anything.
My background runs through computer science, telecommunications, and bioengineering, and that mix shapes how I judge a model. I care far more about whether a result survives on new data than whether it looks impressive on the data it was trained on. I keep my code public and I write down what did not work, because the failures are usually the useful part.
Bioengineering 12(7):699, MDPI
Stacked tree-based models on multimodal voice features. The same model both detects Parkinson's and estimates motor severity (UPDRS), validated with 10-fold cross-validation.
Bioengineering 11(12):1218, MDPI
A Vision Transformer paired with KNN on spiral drawings. The result I did not expect: the model did best on un-augmented data. Augmentation has to be matched to the architecture, not assumed to help.
Chrome Web Store, Manifest V3, OpenAlex API
A browser extension for keeping up with the literature. It saves papers from 10+ publishers, tracks submission deadlines, and suggests journals based on what you work on. It runs entirely in the browser, with no server and no tracking. I built it because I wanted it for myself.
I hid a waveform in the machine. Turn the three dials until your pink trace sits on top of it, then lock it in. This is roughly 4% of my actual job, except the real version has no score and takes months.
The best tools come from being annoyed often enough. On scratching your own itch, and why researchers should write more software.
Reproducibility is not about virtue. It is about the fact that in six months you will not remember anything, and a reviewer will ask.
Moving from Algeria to Tianjin, doing research in English, living in Mandarin, and thinking in French when nobody is looking.