Divan, a Turkish LLM arena, is now live on Spaces. It puts two models side by side on the same Turkish prompt so that differences in reasoning, fluency and instruction-following become directly comparable instead of anecdotal. The space has two modes: an open comparison where you pick both models, and a blind mode that hides model identities until after you have read both answers — removing the brand bias that makes informal LLM comparisons unreliable. Prompts are organized into 6 categories (general knowledge, reasoning, creative writing, summarization, code, translation) with ready-made Turkish test items, and every model runs through the HF Inference API.
CV Matching Engine is now live on Spaces. It takes a PDF CV and a job description and produces an explainable relevance ranking rather than an opaque score: the pipeline runs semantic NLP over the parsed document, applies hybrid scoring that combines semantic similarity with lexical signals, and surfaces the reasoning behind each ranking decision. The goal is to make CV screening auditable — recruiters and candidates can see which parts of a profile drove the match, instead of trusting a single number. Built for Turkish and English CVs.
Mizan, a RAG-based Turkish disinformation fact-checker, is now live on Spaces. It is built on the Directorate of Communications (İletişim Başkanlığı) DMM dataset of 2,810 officially verified disinformation claims, released openly under CC-BY-4.0. Users enter a suspicious claim and the system retrieves the most semantically similar bulletin records, combining Turkish Sentence-BERT embeddings with a FAISS index for millisecond-level search and Qwen2.5-7B for grounded assessment. The output is a three-part verdict: a similarity-scored match indicator, the matched bulletins with their official corrections, and an LLM explanation constrained to the retrieved evidence.
Çarşı, a dual-model Turkish e-commerce NER demo, is now live on Spaces. It runs Named Entity Recognition side by side with a fine-tuned Turkish BERT and Qwen2.5-7B zero-shot extraction, making the "specialized NER vs. zero-shot LLM" trade-off visible in real time. The system extracts 7 entity types from Turkish product text — PRODUCT, BRAND, PRICE, COLOR, SIZE, MATERIAL and GENDER — with color-coded highlighting over the original sentence. The fine-tuned BERT returns entities in ~10-50ms on CPU, while the LLM handles open-ended extraction through the Inference API. Trained on 8,000 synthetic Turkish e-commerce sentences with BIO-tagged token classification and no manual annotation.
Şifahane, a dual-inference medical classification demo, is now live on Spaces. It features side-by-side Turkish BERT and Qwen2.5 architectures for real-time evaluation of the "Classifier vs. LLM" trade-offs, all within a single space. The system utilizes a fine-tuned Turkish BERT for high-speed, cost-effective inference and the Qwen2.5-7B model for flexible multi-task reasoning, with support for department classification, condition analysis, urgency assessment, and rationale generation across 12 medical departments.
