AI researcher — language model evaluation, biomedical AI, and computational biology.
High school student in Istanbul, Türkiye. Co-founder and CTO at Ethosoft, and AI research intern at the Istanbul Technical University UZDES Center, operated in collaboration with the Turkish Space Agency (TUA), where I work on space support systems for the National Space Program.
Most of my research asks a single question in different forms: how much of a model's reported performance is real? I build preregistered, reproducible evaluation protocols that separate genuine capability from experimental artifacts — training-order effects, adaptive compute allocation, and counterfactual verification of safety mechanisms. I publish negative results with the same care as positive ones.
Alongside that, I develop clinical decision support systems from biomedical signals and images (EEG, ECG, EMG, chest radiography, endoscopy, OCT/fundus, echocardiography), and work on protein structure modeling and computational drug screening.
Peer-reviewed
Under review — IEEE Transactions on Dependable and Secure Computing When Safety Is Not Load-Bearing: Counterfactual Dependability Evaluation of Mandatory-Path Safety Mechanisms in Language Models
Preprints
- When Training Order Becomes the Algorithm — endpoint-induced collapse and memorization without rule generalization
- Testing Surprisal as a Compute-Allocation Signal — a preregistered long-horizon study of causal token routing
- NEDOQwen — diagnosing and repairing a Turkish-centric 0.824B language model
- Foresight Is Not Enough — sentence-level future signals and the calibration gap in small language models
Full list on ORCID.
EpiLang — Reproducible codebase on learning and executing the semantics of synthetic languages. Generators, interpreters, immutable experiment configs, a written preregistration, scientific audit reports, 88 test files, three Apptainer definitions, and Slurm recipes for HPC runs. Apache-2.0
surprise-ca — Does causal surprisal make a good signal for allocating extra compute? Preregistered, multi-seed, with FLOP accounting and scaling proxies. The sealed final test was never opened — the answer was no, and the repo says so up front.
mycelium — Public-safe codebase for testing whether a language model safety mechanism is behaviorally load-bearing. Mandatory-path wrapper with counterfactual harness: removal, misrouting, family zeroing, base-path bypass. Raw prompts, generations, and weights deliberately excluded under responsible-release policy. MIT
kuantum — Collision event classifier and detector simulation toolkit. Monte Carlo event generation, GPU-accelerated particle transport, GDML/ROOT geometry import, PyVista 3D rendering, and Three.js/WebXR export.
dockingdemo — Pharmacological chaperone search for the G6PD R166H mutation. AlphaFold wild-type and mutant structures, full GNINA screening results for 3,799 compounds, candidate poses, and a Flask viewer.
- EU Code Week Hackathon — 1st in Europe and Türkiye, Jury Special Award, for a smartphone-integrated portable fundus diagnostic device
- TEKNOFEST International AI in Health — 3rd place, high school category, with cardiomegaly detection (91.4% F1, 96% AUC) and colonoscopy polyp segmentation (0.8907 mean IoU)
- TÜBİTAK 2204-A — national research project approval for AI-assisted drug simulation targeting G6PD mutations
Python C/C++ CUDA Dart
PyTorch JAX TensorFlow Transformers
Linux Slurm/HPC Docker Apptainer Git
AlphaFold GNINA RDKit OpenCV Flutter
Open to research collaborations. Reach me at nedim@ethosoft.org.