Research
Work published with Zirak AI and its researchers, from Pashto language resources to vision-language benchmarks and trustworthy generation.
A clinical dataset of 8,775 dental checkups from 4,800 patients, with 50,000 intraoral photographs, 8,056 radiographs, and written diagnoses. Fine-tuned Qwen-VL models are tested on anomaly classification and full diagnostic reports, against their base models and GPT-4o.
A systematic review identifies sixteen factors that stall global software outsourcing. SWOT, Pareto, clustering, and exploratory factor analysis then group them into a smaller set of constructs teams can actually act on.
An explainable score for fluent but false model output. Token-level SHAP and LIME attributions are turned into a hallucination score and tested on TruthfulQA and QAGS with GPT-3.5, LLaMA-2-13B, and Falcon-40B.
Introduces one million synthetic Pashto OCR images across 1,000 font families, plus a 10,000-image benchmark. Zero-shot tests of Llama, Florence, Qwen, GPT-4o, Gemini, Claude, and Grok show Gemini leading overall and Qwen-7B leading the open models.
A Pashto corpus of about 700,000 words and 30,000 sentences, labeled for word boundaries and 36 part-of-speech tags. A fine-tuned multilingual BERT tagger reaches 96.24% accuracy, ahead of a BiLSTM-CRF baseline.
Fake-news detection that looks past the article itself. Graph networks of user interaction and semantic spread, enriched with psycholinguistic features and BERT embeddings, are classified with graph convolution and multi-head attention.
POLD, a manually labeled Twitter dataset of offensive and non-offensive Pashto. CNN and RNN baselines are compared with fine-tuned XLM-R and a Pashto BERT trained from scratch, which reaches 94.77% accuracy.
Pashto does not mark word boundaries reliably with spaces. Two conditional random field models, trained on a 3.5-million-word corpus, correct space insertion and omission and then segment the text, with F1 scores of 99.2% and 96.7%.
An open toolkit for a language spoken by more than 50 million people: spelling correction, word segmentation, part-of-speech tagging, offensive-language detection, static embeddings, and the first monolingual Pashto BERT.
A classifier that flags fake-account names in Pashto before any social history exists. Trained on 100,000 labeled names and strings, Naïve Bayes is the strongest of the models tested, at 95.3% accuracy.
A permissioned blockchain for drug traceability, from manufacture to the patient, so manufacturers and regulators can see a supply chain that today is fragmented, and so recalls and follow-up are possible.
