Rabiul Alam, Md. GolamReza, Md TanzimNabil, Jonayed KaderChowdhury, Tasnim AzizHarun, Sayed Ilham AzharMozumder, Rafi EhteshamTasnim, Nuzhat2026-04-262026-04-2620262026-01ID 22101100ID 22101164ID 22101262ID 22301489ID 22101167http://hdl.handle.net/10361/28064Cataloged from PDF version of thesis.Includes bibliographical references (pages 47-50).This thesis is submitted in partial fulfillment of the requirements for the degree of Bachelor of Science in Computer Science and Engineering, 2026.Recorded class lectures and tutorials are essential learning resources, but unedited raw videos generally contain significant non-instructional content, such as silence, administrative discussions, off-topic instructions, making effective navigation and revision difficult. This study presents a framework which is open-source, contextaware and can automatically detect and remove irrelevant or off-topic segments from long, unedited educational videos. Most of the existing related solutions rely heavily on large, closed-source models or fixed-length video segmentation with no context awareness, which are computationally expensive and prone to breaking semantic continuity. To address these limitations, we propose a modular, four-stage orchestration pipeline designed to run entirely on consumer-grade GPUs using small, open-weight models. The framework uses automatic speech recognition (Parakeet TDT 0.6B) to perform dynamic, sentence-level video segmentation, ensuring that each segment represents a complete semantic unit. A lightweight vision language model then generates structured textual descriptions for each segment using both visual frames and subtitles, augmented with rolling local context to preserve chronological coherence. Finally, a small language model (Qwen3-4B-instruct) is used to perform classification on the generated descriptions rather than on raw video input. The proposed orchestrator pipeline achieves 96.85% accuracy, 89.04% precision, 92.77% recall, and 90.86% F1 for irrelevant-segment detection on a human-annotated benchmark of 20 educational videos (total duration 09:13:47). This results in a significant improvement in content preservation over a one-shot Gemini-3 baseline, with precision increasing from 45.98% to 89.04% and accuracy from 81.63% to 96.85%. The entire system runs locally via quantized inference, offering a practical and privacypreserving alternative to cloud-based solutions.53 pagesenBRAC 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.Educational video analysisIrrelevant content detectionVision-language modelsSpeech recognitionVideo processingMachine learningEducational technology.Computer-assisted instruction.Educational technology.Speech processing systems.Thesis