PORTFOLIO

The rapid expansion of scientific literature has made it increasingly difficult for researchers to extract meaningful insights from dense academic documents. We introduce LongDocAI, a multimodal, multi-model framework designed to efficiently understand and interact with scholarly PDFs at scale. The system combines Donut, an OCR-free document parser, with a hybrid summarization module based on BARTlarge- CNN, trained respectively on over 200,000 paper-summary pairs and 30,000+ question-answer examples.

To ensure faster and resourceefficient inference, we apply LLM.int8 post-training quantization across all transformer models, significantly reducing memory usage and latency without compromising output quality. LongDocAI is built as a unified pipeline that handles both layout-aware parsing and semantic-level understanding through summarization and interactive question answering.

Results: 15% improvement in ROUGEL, 28% boost in METEOR, 85.3% QA accuracy. 70% reduction in model size and 40% decrease in latency.

This paper presents the design, quantization, and deployment of an embedded Advanced Driver Assistance System (ADAS) prototype on a Raspberry Pi 3, utilizing a custom-trained YOLOv12n object detection model. The system achieves real-time multi-class object detection and autonomous control of a test vehicle within the resource constraints of low-power edge devices.

To optimize performance, posttraining INT8 quantization was applied to YOLOv12n, resulting in over 65% reduction in model size and nearly 3× faster inference, with minimal accuracy loss. Object detection outputs are combined with ultrasonic sensor measurements to enable obstacle-aware vehicle control, including braking and navigation.

Results: 6 FPS on RPi 3, 210ms control latency, >65% reduction in model size.

This paper introduces YOLO-OptiMob , a comprehensive pipeline for optimizing YOLO11 models for deployment on edge devices. The process begins with creating and preprocessing a custom dataset featuring seven object classes. The YOLO11 model is trained on this dataset and optimized using L1 unstructured pruning, with rates of 30%, 40%, and 50% evaluated.

Based on the results, a 40% pruning rate was selected as it offered the best balance between model size reduction and accuracy retention. Post-training INT8 quantization further compresses the model, reducing its size from 11.4 MB to 2.5 MB. The optimized model is then converted into TensorFlow Lite (TFLite) format, ensuring compatibility with Android-based edge devices.

Results: Reduced model size from 11.4 MB to 2.5 MB while retaining high detection accuracy.