Online recruitment scams pose a growing threat in digital job markets, particularly in emerging economies such as Bangladesh, where socioeconomic conditions create a highly vulnerable population. With a youth unemployment rate of 7.2%, a gender skewed victimization pattern (59% of the victims are female), and widespread adoption of mobile financial services such as bKash and Nagad, Bangladesh presents a uniquely challenging environment for fraud detection. Existing machine learning models for scam detection have been predominantly trained on Western datasets and therefore fail to capture fraud signals specific to emerging markets. This paper proposes a localized detection framework based on an Enhanced Multilingual BERT architecture with custom regularization, trained on a dataset of 20,000 annotated job postings collected from major Bangladeshi recruitment platforms (BDJobs and LinkedIn) as well as informal channels such as Telegram and WhatsApp groups over a 21-month period. The dataset incorporates Bangladesh-specific fraud signals, including references to mobile financial services, informal communication patterns and Bengali-English code-mixing. The proposed model achieves a test accuracy of 95.6% with precision of 94.0% and 89.0% for legitimate and fraudulent postings respectively and F1-scores of 95.0% and 87.0% for the same classes. The framework outperforms existing global models on Bangladesh-specific evaluation data by 3.3 to 8.1 percentage points.