Neuro-physiotherapist and neuroscience researcher — mechanistic interpretability, lesion and compensation in neural networks.

I study how neural systems, artificial and biological, lose and recover function. After more than eight years of clinical practice in neurorehabilitation, I now carry the clinical logic of lesion, diaschisis and compensation into large language models: I lesion circuits, map how the rest of the network responds, and test whether activation steering can partially restore what was lost. The throughline is a single aim — turning opaque systems into ones whose behaviour can be explained, predicted and reasoned about.

MSc thesis, Neuroscience — Ege University (in progress) Functional Compensation After Induced Neural Lesions in Large Language Models: Testing Partial Rehabilitation through Causal Circuit Re-mapping and Activation Steering.

• Question: when a causally identified circuit is ablated, which components take over, and can that compensation be summarised in a single, interpretable coefficient?
• Method: GPT-2 Small/Medium and the Pythia family; subject–verb number agreement as the main task, IOI for method validation; activation patching, ablation as a digital lesion, pre- vs post-lesion causal maps, and dose–response activation steering.
• Measures: logit difference, recovery rate, a specificity check on unrelated text, and a compensation index (ERI) adapted from the Hub Disruption Index, with permutation tests and bootstrap intervals.
• Planned outputs: an open-source code repository and a workshop paper or arXiv preprint.

• Mechanistic interpretability — circuits, causal interventions (activation patching, ablation, steering), and how networks compensate for damage (self-repair, the Hydra effect), read through systems neuroscience.
• Neurorehabilitation and computation — translating clinical neuro-physiotherapy into computational models; neuroimaging-based biomarkers and clinical outcome prediction.

• Clinical: paediatric and adult neurorehabilitation across a tertiary hospital, special-education and rehabilitation centres, a disability care centre and independent practice.
• Education: MSc Neuroscience, Ege University (thesis stage) · BSc Physiotherapy and Rehabilitation, Dokuz Eylül University · two years of Electrical and Electronics Engineering, Gaziantep University.

• Self-hosted, encrypted research infrastructure — a declarative, reproducible environment on NixOS with full-disk encryption, a private WireGuard mesh with key-only SSH, self-hosted file synchronisation, and encrypted, versioned off-site backups.
• Helix Epistemica — a browser-based instrument for calibrated belief updating: it computes evidence thresholds and credence changes, flags reasoning pathologies, and exports to a Git-versioned note vault.

Remote work in AI evaluation and clinical or rehabilitation domain expertise, and research collaboration in mechanistic interpretability. English C1 (TOEFL iBT 102) · Turkish native.