Blog

Why TF-IDF struggles with little data, and how language-model embeddings with logistic regression lifted an off-topic filter from 0.69 to 0.82.

Fine-tuning XLM-RoBERTa on Turkish and English financial headlines: leakage control, weighted loss, training setup and 0.79 macro-F1.

Lessons from three live projects on Filament v3, v4 and v5: when it fits, when it does not, actions, authorisation and version changes.

What a Turkish law firm website may and may not include under the TBB advertising regulation: references, keywords, ads and the 2026 centre.

Calm design, expert content, privacy-safe contact and the 2025 Turkish health promotion rules on testimonials, fees and SEO.

A model scoring 0.79 on the benchmark dropped to 0.53 on real Turkish headlines. The cause was domain shift; a neutral decision threshold lifted it to 0.72.

Instead of splitting the corporate site, admin panel and client portal apart, I kept them in one project — drawing the boundary with route groups, query scopes and file-access rules rather than folders.