AI Engineer exploring different world

devto

Nouman Amjad  • Peerlist

Benjamin Franklin was the king of "Poor Richard’s Almanack" and colonial humor. I built <a target="_blank" rel="noopener noreferrer" class="break-all link-break text-gray-500 hover:underline" href="http://benmode.com">benmode.com</a> so you can trade jokes or debate philosophy with him in real-time.Whether you need a financial tip, a moral compass or a solution to a modern problem, get your respond from the Founding Father of the USA in his prespective.It’s in Beta, try to break the bot or find its funniest response and share it here! Voice Mode coming soon if you guys like it.Visit <a target="_blank" rel="noopener noreferrer" class="break-all link-break text-gray-500 hover:underline" href="http://Benmode.com">Benmode.com</a> Now!

Benmode

Developed an AI-driven attendance system using YOLOv8 & FaceNet, integrating a Vector Database ChromaDB for efficient face embedding storage and retrieval. Eliminated the need for retraining by leveraging retrieval-augmented generation (RAG)-style lookups, reducing computational overhead and improving accuracy. 1. Face Embedding Storage: Used ChromaDB to store and retrieve face embeddings, enabling instant recognition without retraining. 2. Real-Time Attendance Marking: Captured classroom images and matched them against stored embeddings for automated search, filter, and verification. 3. FastAPI Deployment: Deployed the system using FastAPI, ensuring high-speed inference and scalable API integration for seamless attendance tracking. 4. Scalable Web Application: Built with FastAPI & Bootstrap, providing student management, attendance history, and cross-device accessibility. 5. Enhanced RAG-based Face Retrieval: Improved system efficiency using approximate nearest neighbor (ANN) search, enabling fast, scalable face recognition.

Automated Attendance Tracking using YOLOv8 & 

Developed an AI-powered georeferencing system to map UAV images to satellite imagery, enabling precise geo-coordinate retrieval without using GPS. Integrated local feature extraction techniques with homography estimation for accurate spatial alignment. 1. Local Feature Extraction: Implemented SIFT & SURF for robust keypoint detection and feature matching. 2. Homography Estimation: Utilized RANSAC to filter outliers and compute accurate image transformations. 3. Real-time Georeferencing: Mapped UAV image pixels to satellite coordinates using pixel projection and interpolation. 4. Scalable & Efficient: Optimized feature matching for large-scale aerial datasets, improving spatial accuracy and computational performance.

Automatic Georeferencing System using Local F

Developed an advanced event management portal featuring real-time notifications, user profiles, and feedback mechanisms. Ensured cross-device compatibility and a responsive design to enhance user interaction and event accessibility.

Events Management System

Contributed to a cutting-edge R&D project focused on geotagging US roadway assets using 2D images. Leveraged deep learning techniques such as depth estimation models and advanced geolocation methods to achieve geotagging accuracy under seven feet, significantly surpassing existing systems. 1. Fine-tuned depth estimation models and processed imagery for diverse weather conditions. 2. Conducted extensive research to overcome the absence of LiDAR depth maps by implementing innovative algorithmic adjustments. 3. Demonstrated technical expertise in Python, TensorFlow, and PyTorch while testing and optimizing the system for scalability and precision.

Geotagging US Roadway Assets using 2D Images

Created a secure matchmaking platform that prioritizes user values and beliefs while maintaining privacy. Integrated user-friendly features to ensure seamless connections while adhering to ethical principles.

Halal Pairing: Matchmaking App

Developed a real-time environmental monitoring and pollution prediction system using modern MLOps tools. The project included automated data collection with the OpenWeatherMap API, a predictive LSTM model deployed with Flask, and real-time system monitoring using Prometheus and Grafana. Key features include: 1. Data version control with DVC and model tracking with MLflow. 2. Deployment of the trained LSTM model for real-time predictions. 3. Advanced monitoring and visualization with Prometheus and Grafana.

Pollution Prediction System With MLflow, Prom

Designed a scalable and reproducible DVC pipeline to train and evaluate a Random Forest Classifier. Key components of the pipeline included: 1. Automated stages for data preparation, feature extraction, model training, and evaluation. 2. Reproducible results using version-controlled data and models. 3. Evaluation metrics with visualization of feature importance and ROC curves. 4. Configured pipeline parameters using dvc.yaml and params.yaml for easy tuning

Random Forest Classifier With DVC Pipeline

Developed a synthetic data generation pipeline to improve object detection accuracy for roadway assets by 60%. Utilized GANs, Transformers, and Diffusion Models to create thousands of synthetic images, augmenting the KITTI dataset and achieving significant improvements in precision (mAP) metrics.

Synthetic Data Generation for Enhanced Object

Built a responsive travel agency website with automated search and booking functionality. Integrated real-time search features, customizable filters, and enhanced user experience with modern web frameworks.