Identification of peptide immunogens in a pathogen proteome can help to develop diagnostic kits, therapeutics, and preventive vaccines. Traditional methods for identifying such peptides are often laborious, time-consuming, and resource intensive, resulting in limited efficiency and scalability. Given the extensive size and diversity of viral proteomes, the use of state-of-the-art computational tools has become essential for identifying immunogenic fragments. This study presents MHCmet, a deep learning tool for identifying such peptide fragments that bind human MHC class I receptors. The key innovation of MHCmet lies in its allele-specific encoding strategy, integrating with CNN-bidirectional LSTM layers in the architecture. Whereas existing tools apply a single encoding scheme uniformly across all HLA alleles, MHCmet dynamically selects the optimal encoding technique One-hot, BLOSUM62, or Non-Linear Fisher (NLF) for each allele based on predictive performance with the integration of Convolutional Neural Networks (CNN), multi-head self-attention, and bidirectional LSTM layers. This design captures the structural and physicochemical diversity of peptide-MHC binding grooves better, leading to measurably improved predictions. MHCmet had a median AUC of 0.91, higher than NetMHCpan (0.89) and MHCflurry (0.85). The tool accepts target protein sequences in FASTA format as input and employs various encoding techniques depending on the MHC class I supertype. The output file contains the epitope sequences, epitope lengths, binding affinities, prediction scores, and the MHC that binds each epitope. Furthermore, the tool was used to predict MHC class-I epitopes, targeting the glycoprotein of Dobrava Orthohantavirus (DOBV) and identified more than 40 peptide fragments with binding affinity for MHC class-I receptors. These candidates require further experimental validation before any translational application. The tool can be freely accessed at https://github.com/rishabhdhenkawat/mhcMET.