Ph.D. Candidate · Data Scientist & AI Researcher
Most models are evaluated on data that looks like their training set. Deployed systems rarely get that luxury. Sensors drift, hardware differs, environments change, and the cost of a confident wrong answer can be someone's diagnosis or someone's location under rubble.
My work sits at that failure point: making models robust when the data distribution moves, explainable when the stakes are high, and small enough to run where the data is actually collected.
Istanbul, Türkiye
My publications look like they span medical imaging, wireless sensing, and video quality. They don't. They are the same problem seen from three angles: what happens to a model when reality stops matching the training set?
If I train on one device today, why does it fail on another device tomorrow?
Hardware heterogeneity and temporal drift silently violate the i.i.d. assumption that nearly every deep architecture rests on. I work on neutralizing that divergence at the signal level, through statistical alignment and physics-aware preprocessing, rather than papering over it with more parameters.
Would a clinician stake a decision on this prediction?
Accuracy alone is not evidence of understanding. In medical imaging and safety-critical sensing I pair predictions with gradient-based attribution, so a domain expert can interrogate why the model decided, and catch the cases where it is right for the wrong reason.
Does it still work on the hardware that actually exists?
A model that needs a datacenter is useless on a rescue robot or a low-cost IoT node. I profile latency as a first-class result, use sparse Mixture-of-Experts routing to buy capacity without paying for it at inference, and treat deployment as part of the science.
Each one written the way I'd present it to a research group: the question, the prior art it breaks, the method, the numbers, and what didn't work.
Wi-Fi CSI fingerprinting collapses when you test it on a different day than you trained it. Instead of discarding corrupted packets, this framework restores the physical signal, recovering saturated receivers and formally re-establishing the i.i.d. assumption across temporally disjoint domains.
CNNs see texture; they struggle with the global context radiologists rely on. Pairing convolutional locality with transformer attention closed that gap, and Grad-CAM attribution turned the output into something a clinician can actually audit.
Two receivers in the same spot report different RSSI. Deep feature learning over the TUJI1 dataset recovers a device-invariant representation, so a fingerprint collected on one handset still means something on another.
This is not a simulation. Below are two Wi-Fi CSI captures I recorded at the same location, on the same radio and channel, 19.4 minutes apart. Nothing moved but the clock, yet the measurements drifted enough that a classifier can tell the two sessions apart almost perfectly. Everything is computed live in your browser.
Reference: ref_204_1_3500.pcap (3500 packets) · Test: test_204_1_100.pcap (100 packets) · 52 usable subcarriers · PCA basis fitted on the reference session only.
How to read this: the session discriminator is a logistic classifier trying to guess which capture a packet came from. 50% means it cannot tell: the sessions are statistically indistinguishable, which is what we want. High accuracy means a localizer trained here would learn when the data was captured rather than where the device is.
CORAL whitens the test session and re-colours it with the reference covariance, X′ = (X−μt)·Σt−1/2·Σs1/2 + μs, using no labels from the test session.
Research earns its keep when it survives a deadline and a real user. These are production systems: running, maintained, and answerable to someone other than a reviewer.
A trade-show lead pipeline that turns a pile of business cards into clean, queryable records. Gemini Flash 2.5 reads each card directly, and the SQLAlchemy/PostgreSQL schema resolves duplicates on ingest rather than during a later cleanup.
A donation platform for orphan families in Sudan, built so donors never have to take the word "distributed" on trust. Every transfer is verified, converted at a recorded rate and disbursed against an official receipt, with per-family timelines published openly.
A corporate platform for a medium- and low-voltage switchgear manufacturer, built around one route: unfamiliar visitor to qualified quote request. Parallel Turkish/English routing, global search, and KVKK compliance handled up front.
Each with a plain-language note on why it matters.
Telecommunication Systems (Springer), 2026
IEEE Access, 2025 · doi:10.1109/ACCESS.2025.3621505
@article{saeed2025mitigating,
author = {Saeed, M. T. M. and Yousif, M. A. A. and Ozturk, I.},
title = {Mitigating Device Heterogeneity for Enhanced Indoor Positioning
System Performance Using Deep Feature Learning},
journal = {IEEE Access},
year = {2025},
doi = {10.1109/ACCESS.2025.3621505}
}
Electronics (MDPI), 2026, 15(8), 1666 · doi:10.3390/electronics15081666
@article{dutta2026impact,
author = {Dutta, A. and Saeed, M. T. M. and Arawade, S. and Sam{\v{c}}ovi{\'c}, A.
and Uddin, S. and Juszka, D. and Grega, M. and Leszczuk, M.},
title = {Impact of Environmental Control on Subjective Video Quality
Assessment in Crowdsourced QoE Experiments},
journal = {Electronics},
volume = {15},
number = {8},
pages = {1666},
year = {2026},
doi = {10.3390/electronics15081666}
}
2024 12th Int. Scientific Conference on Computer Science (COMSCI), Sozopol, Bulgaria · doi:10.1109/COMSCI63166.2024.10778503
@inproceedings{saeed2024transconvnet,
author = {Saeed, M. T. M. and Yousif, M. A. A. and Hakim, A. A. M.},
title = {TransConvNet: Enhancing Kidney Abnormality Detection in CT Imaging
through Hybrid Transformer-CNN Model with Integrated Explainability},
booktitle = {2024 12th International Scientific Conference on Computer Science (COMSCI)},
address = {Sozopol, Bulgaria},
pages = {1--5},
year = {2024},
doi = {10.1109/COMSCI63166.2024.10778503}
}
IEEE COMSCI 2024, Bulgaria · doi:10.1109/COMSCI63166.2024.10778519
Control engineering taught me to think about systems and feedback; machine learning gave me better function approximators. That instinct (model the physics first, then learn the residual) comes from the former.
AGH University of Science and Technology · Kraków, Poland
TÜBİTAK & Osmaniye Korkut Ata University · Türkiye
AGH University of Science and Technology · Kraków, Poland
Al Neelain University · Khartoum, Sudan
Amsterdam, Netherlands
Split, Croatia · sponsored by Google DeepMind
Fully funded by Mastercard
Ondokuz Mayis University · Samsun
Istanbul University-Cerrahpaşa
Al Neelain University · Khartoum
I'm interested in roles and collaborations where the data is messy, the deployment is real, and the failure modes matter. If that sounds like your problem, get in touch.