Chakrabarty, AmitabhaBhoumik, ParthaArjan, Promit Dey SarkerMati, Mrittika DeviBasak, ArghaBari, Saib Sadman2026-08-102026-08-1020262026-01ID 24141134ID 24341241ID 22101398ID 22101668https://hdl.handle.net/10361/28852This thesis is submitted in partial fulfillment of the requirements for the degree of Bachelor of Science in Computer Science and Engineering, 2026.Cataloged from PDF version of thesis.Includes bibliographical references (pages 87-90).Agriculture plays a vital role in Bangladesh, employing approximately 43% of the population, yet many farmers lack formal education and expert guidance, leading to uninformed decision-making. This research explores the development of a multimodal Visual Question Answering (VQA)-based agricultural assistance system that integrates computer vision, natural language processing (NLP), and speech processing to provide real-time, voice-based responses in Bangla, reducing the immediate need for agricultural experts. Instead of relying solely on CNN-based classification, the system employs Vision–Language Models (VLMs) with visual encoders for crop disease understanding, a Large Language Model (LLM) for agricultural query answering, and speech models for Bangla voice interaction. The system offers an intuitive, offline-compatible mobile solution tailored for farmers in low-resource environments. A conversational agricultural diagnostic dataset was developed containing 96,003 expert-style diagnostic entries with 9,138 rice leaf images across six classes. A structured dataset covering Bangladesh’s diverse crops, pests, and diseases ensures AI model optimization and practical usability. A comparative evaluation assesses the multimodal AI system (Image + Text + Voice) against traditional single-input models (Image-only, Text-only) based on diagnostic accuracy, response relevance, and user satisfaction. In the final offline pipeline, MobileNetV3-Small achieved 95.35% disease classification accuracy with stable mobile CPU latency ( 12– 18 ms), Gemma 3–1B produced the strongest overall response quality among tested lightweight LLMs (METEOR = 0.1095, cosine similarity = 0.539, and BERTScore F1 0.70, BertScore Precision: 0.72, BertScore Recall: 0.69), and Whisper-Base enabled faster-than-real-time Bangla speech recognition (RTF = 0.7859, WER = 0.353). This approach minimizes dependence on agricultural experts, saving time and costs while positively impacting the economy. A farmer field survey (n = 30) using five evaluation parameters indicates strong real-world usability: 92% rated the system helpful/very helpful, 88% found it easy to use, 84% reported successful offline usage, 86% expressed satisfaction with response clarity, and 82% reported improved confidence in decision-making. Ultimately, this research seeks to bridge the digital divide in agriculture, empowering farmers with an AI-driven solution for real-time problem-solving and enhanced food security.102 pagesen-USAttribution-NonCommercial-NoDerivatives 4.0 InternationalBRAC University theses are protected by copyright. They may be viewed from this source for any purpose, but reproduction or distribution in any format is prohibited without written permission.http://creativecommons.org/licenses/by-nc-nd/4.0/Natural language processingComputer visionSpeech processingLow-resource environmentsMultimodal AIVisual question answeringDiseases detectionCrop diseasesAgricultural assistance systemLarge language modelsVision language modelsBengali languageArtificial intelligence--Agricultural applications.Natural language processing (Computer science).Plant diseases--Diagnosis.Thesis