Human action classification system with pose-based (MediaPipe) and video-based (3D CNN) models. Features 100+ architectures for real-time pose classification and temporal models pretrained on UCF-101/HMDB51. Applications include autonomous vehicles, video surveillance, and action recognition research.

Securade.ai Sentinel - A monitoring and surveillance application that enables visual Q&A and video captioning for existing CCTV cameras.

A real-time inferencing of multistreaming YOWOv3(Spatio Temporal Action Detection task) using (UCF101-24) dataset. The repo is extension of https://github.com/Hope1337/YOWOv3, https://arxiv.org/pdf/2408.02623

YOLO-TLP: detected and classified tiny objects with bounding box dimensions smaller than 15 pixels, outperforming other one-stage detectors. maximum resolution for target observation in real-time applications.

Custom YOLO11m model for detecting and classifying car body damage (99% shattered glass, 96% flat tire detection accuracy)—optimized for high-capacity inference and assistive use in inspection and service workflows like BMW pre-loaner inspections.

This project uses YOLO for real-time leukemia detection in blood samples and CNNs for classifying brain hemorrhages in MRI scans. It aims to support faster, more accurate medical diagnostics through deep learning.

Explore a wide range of computer vision projects and documentation covering everything from object detection, image segmentation, and tracking to pose estimation, object counting, and automated annotation. These resources highlight real-world AI applications built with modern models like Ultralytics YOLO, Meta SAM 2, and other vlms

YOLOv12 Underwater Object Detection is an open-source suite for underwater object detection, built on YOLOv12. It offers an end-to-end pipeline with GPU-accelerated training, customizable data augmentations, real-time inference via Gradio, and support for model export (ONNX & PyTorch).

This project uses YOLO for real-time leukemia detection in blood samples and CNNs for classifying brain hemorrhages in MRI scans. It aims to support faster, more accurate medical diagnostics through deep learning.

Improved PIDNet for real-time semantic segmentation. Work in progress.

THYX is an edge AI video analysis platform for AIoT, featuring behavior recognition, intelligent alerts, and real-time management. Modular, high-performance, and lightweight — ideal for smart security and industrial scenarios.

A conveyor belt sorting system powered by Raspberry Pi and YOLOv8 for real-time object detection.

Volleyball tracking - VballNet is a specialized deep learning framework designed for volleyball tracking, built upon the foundation of TrackNetV4. This repository includes two primary models, VballNetV1 and VballNetFastV1

Project Agora: MVP of the Concordia framework. An ethical, symbiotic AI designed to foster and protect human flourishing.

PlantAi is a ResNet-based CNN model trained on the PlantVillage dataset to classify plant leaf images as healthy or diseased. This repository includes PyTorch training code, tools to convert the model to TensorFlow Lite (TFLite) for deployment, and an Android app integrating the model for real-time leaf disease detection from camera images.

40x faster AI inference: ONNX to TensorRT optimization with FP16/INT8 quantization, multi-GPU support, and deployment

This repository features an image sharpening pipeline using Knowledge Distillation. A high-capacity Restormer acts as the teacher model, while a lightweight Mini-UNet is trained as the student to mimic its performance.

Bachelor Final Year Project exploring real-time speech denoising using machine learning. Compares classical methods (SS, WF, MMSE-LSA) with 5 deep models on spectrogram data, highlighting Conv-TasNet’s effectiveness. Features dataset bucketing, OOM mitigation, and batch evaluation.

LLM-inspired BiLSTM pipeline for real-time, multi-label toxicity inference across adversarial discourse modalities.

Deep learning system for real-time face mask detection achieving 98.2% accuracy using ResNet18 architecture, PyTorch, and Gradio. Features efficient inference (0.12s), optimized model size (44.7MB), and a balanced dataset of 20,000 images for robust production deployment.