Meet
Sharoze Ali
Lead AI Engineer & Computer Vision Expert specializing in smart cities and industrial automation
7+ years transforming AI research into real-world solutions • 70% cost cutting & accident reduction • 40+ projects deployed
Contact
Featured Projects
A selection of AI solutions built for scale, safety, and real-world impact.
Satellite Building Analysis
Accurately classify urban structures at scale.
Industrial Safety AI
Reduced workplace accidents by 70% with real-time anomaly detection.
Metro Crowd Intelligence
Deployed in 40+ metro stations to optimize passenger flow.
Heritage Recognition
Advanced AI for preserving cultural landmarks.
Proven Impact
7+
Years
of experience
70%
Accident
reduction
40+
Projects
deployed
AI Passion & Vision
My Mission
I have a great passion to enhance technology with AI. My interests span robotics, computer vision, and 3D reconstruction where I apply machine learning algorithms to train and automate intelligent systems.
I specialize in image recognition, multi-object tracking algorithms, and deep learning model optimizations for onboard processing boards like NVIDIA Jetson TX2 Nano and AGX/NX Xavier.
Career Goals
I want to be part of any firm that develops my skills to excel in the AI market and is well-known for contributing to society as an IT industry leader.
My focus is on transforming AI research into real-world solutions for smarter cities and safer workplaces.
Core Technical Expertise
Machine Learning
Deep learning algorithms, model optimization, and edge AI deployment
Computer Vision
Object detection, tracking, segmentation, and real-time processing
LLMs & RAG
Large language models, retrieval systems, and multimodal AI
Edge Computing
NVIDIA Jetson, TensorRT optimization, and embedded systems
Professional Experience at Netcad
3 years 4 months of transformative AI development in Istanbul, Turkey
1
Computer Vision & ML Expert
June 2022 - June 2025
Leading advanced AI projects for smart cities and industrial applications
2
Computer Vision Engineer
March 2022 - June 2023
Specialized in edge AI solutions and urban technology systems
Breakthrough AI Projects
Satellite Building Analysis
Developed satellite-based building age detection using instance segmentation and SAM-GEO framework, helping urban planners identify high-risk zones and boost disaster readiness.
Heritage Recognition System
Built AI-powered heritage recognition system for Ministry of Culture, enabling tourists to instantly identify historic monuments via image input using FAISS and deep embeddings.
Industrial Safety AI
Deployed edge AI safety solution for electric forklifts at KEAS manufacturing plant, reducing accidents by 70% through real-time object detection and red zone alerts.
Featured AI Projects Portfolio
Satellite Building Analysis System
Utilized instance segmentation and the SAM-GEO framework to detect and classify building ages from satellite imagery. This system enhances urban planning by identifying structures vulnerable to natural disasters, significantly improving early warning systems and infrastructure resilience.
Tech Stack: Python, PyTorch, SAM-GEO, OpenCV, AWS S3.
Heritage Recognition AI Platform
Developed an AI-powered platform for instant recognition of historical monuments. Using deep embeddings and FAISS for efficient similarity search, tourists can simply snap a photo to get detailed information, enriching cultural experiences and preserving heritage.
Tech Stack: TensorFlow, Keras, FAISS, REST API, Mobile SDK (Android/iOS).
Industrial Safety Edge AI
Implemented an edge AI solution for electric forklifts at KEAS manufacturing, reducing accidents by 70%. The system provides real-time object detection and "red zone" alerts, preventing collisions and significantly enhancing workplace safety and operational efficiency.
Tech Stack: NVIDIA Jetson, TensorRT, YOLOv8, Custom CNN, MQTT.
Metro Crowd Intelligence System
Engineered a real-time AI system for crowd monitoring and analysis in metro stations. This solution processes video feeds to identify crowd density, bottlenecks, and potential safety hazards, allowing operators to manage passenger flow efficiently and ensure public safety during peak hours.
Tech Stack: OpenCV, DeepStream, Kafka, Prometheus, Grafana.
N-GPT RAG Document System
Designed and deployed an advanced Retrieval-Augmented Generation (RAG) system using Large Language Models (LLMs) for intelligent document processing. This system provides highly accurate and contextually relevant answers from vast document repositories, boosting information retrieval efficiency.
Tech Stack: OpenAI API, LangChain, Pinecone, AWS Lambda, Python.
Volumetric Drone Calculation AI
Developed a drone-based AI system for precise volumetric calculations of stockpiles and construction materials. Leveraging 3D reconstruction techniques and advanced photogrammetry, this project delivers highly accurate inventory measurements for industrial applications, reducing manual effort and errors.
Tech Stack: Pix4D, Metashape, Point Cloud Library (PCL), Python.
Satellite Building Analysis: AI-Powered Urban Risk Assessment
Urban environments are dynamic, and understanding their evolution is critical for effective city planning, infrastructure development, and disaster preparedness. Traditional methods for assessing building characteristics are often labor-intensive, costly, and quickly become outdated. This project addresses these challenges by leveraging advanced AI and satellite imagery to provide a real-time, comprehensive understanding of urban structures.
Project Overview
Empowering urban planners with precise, AI-driven insights into building age and structural vulnerability, facilitating proactive disaster mitigation and sustainable development.
Our "Satellite Building Analysis System" is an innovative solution designed to automate the detection and classification of building ages from high-resolution satellite imagery. By pinpointing older, potentially more vulnerable structures, the system significantly enhances urban resilience and allows authorities to prioritize interventions in high-risk zones, such as those prone to seismic activity or extreme weather events.
Technical Architecture & Methodology
The core of our system relies on a multi-stage AI pipeline, integrating state-of-the-art computer vision techniques. The methodology is centered around robust instance segmentation and a novel approach to infer structural age from visual features and historical data.
  • Satellite Imagery Acquisition: Integration with various satellite data providers (e.g., Planet Labs, Maxar) for multi-spectral and high-resolution imagery.
  • Data Preprocessing Pipeline: Orchestrated via AWS S3 and Lambda for scalable ingestion and radiometric/geometric correction of raw satellite images.
  • Instance Segmentation with SAM-GEO: Employing the Segment Anything Model (SAM) fine-tuned for geospatial data (SAM-GEO) to precisely delineate individual building footprints.
  • Building Age Classification: A dedicated deep learning model (ResNet-50 backbone) trained on a curated dataset of buildings with known construction dates, utilizing features extracted from segmented images.
  • Urban Planning Integration: Outputting georeferenced building polygons with predicted age metadata, compatible with GIS platforms for urban planning and risk assessment.
Implementation Details
Data Preprocessing
Raw satellite images are subjected to a rigorous preprocessing pipeline. This includes atmospheric correction, orthorectification, and normalization. We leverage a custom workflow in Python using GDAL and OpenCV for batch processing, ensuring data consistency and quality for downstream AI models.
Instance Segmentation
The Segment Anything Model (SAM) adapted for geospatial data (SAM-GEO) accurately isolates individual building instances. This provides precise polygon boundaries for each structure, a crucial step for subsequent analysis. For challenging cases, a Mask R-CNN model acts as a fallback or refinement layer.
Building Age Classification
A Convolutional Neural Network (CNN) architecture, specifically a fine-tuned SAM, is employed for acquiring building masks along with geo-coordinates classification. It analyzes visual masks in coming years in the same geo-coordinates, also on the ground building facades, bases , floors, architectural styles (old or new), and patterns within each segmented building footprint. A multi-class classification approach groups buildings into age ranges (e.g., post-2000 to 2023).
Urban Planning Integration
The final output—a GIS-ready layer with building polygons attributed with predicted age—is integrated into city planning systems via GeoJSON APIs and PostgreSQL/PostGIS databases. This allows planners to visualize and query vulnerable zones directly within their existing tools.
Results & Impact
The system demonstrated significant performance gains and real-world applicability during pilot deployments in several municipalities.
92.5%
Segmentation Accuracy
F1-score for building instance segmentation.
88.3%
Age Classification Accuracy
Weighted F1-score across all age classes.
10x
Processing Speed
Faster than manual survey methods.
  • Enhanced Disaster Readiness: Proactive identification of structurally vulnerable buildings allows for targeted reinforcement or evacuation planning, reducing potential casualties and damages during natural disasters.
  • Optimized Resource Allocation: City planners can allocate resources more efficiently for maintenance, renovation, or demolition projects based on empirical data.
  • Sustainable Urban Development: Supports data-driven decisions for urban renewal, preserving heritage sites while modernizing infrastructure.
  • Cost-Effectiveness: Reduces the need for expensive and time-consuming ground surveys, offering a scalable solution for large urban areas.
Technical Challenges Solved
  • Satellite Image Resolution Limitations: Addressed through multi-scale feature extraction in CNNs and integration of higher-resolution commercial imagery where available.
  • Multi-temporal Analysis: Developed techniques to handle variations in satellite image acquisition times and atmospheric conditions, ensuring consistent building feature extraction over time.
  • Ground Truth Validation: Implemented a semi-supervised learning approach, combining limited ground truth data with active learning to continuously improve model performance with minimal manual annotation effort.
Technology Stack
1
Programming
Python (3.9.x)
2
Frameworks
PyTorch (1.13.1), Pytorch, SAM-GEO, Open3D, PyQT, Flask
3
Libraries
OpenCV (4.6.0), GDAL (3.4.3), scikit-image, numpy, pandas
4
Cloud & Databases
AWS S3, AWS Lambda, PostgreSQL/PostGIS (14.x)
5
Deployment
Docker, Kubernetes (for scalable inference)
Future Enhancements & Scalability
  • 3D Building Reconstruction: Integrate with photogrammetry techniques to generate 3D models of buildings for more detailed structural analysis.
  • Damage Assessment Post-Disaster: Develop models for rapid post-disaster damage assessment using pre/post-event satellite imagery.
  • Real-time Monitoring: Explore integration with real-time satellite feeds for continuous urban change detection.
  • Expand Global Coverage: Scale the system to cover additional cities and regions worldwide, leveraging cloud infrastructure and distributed processing.
Code & Architecture Visualizations
While full code samples are proprietary, we utilize version control (Git) and maintain comprehensive documentation. Below is an example of a system architecture component typically showcased in our technical presentations.
For specific algorithm details and model configurations, please refer to our technical whitepapers or request a direct demonstration.
Industrial Safety Edge AI: Real-Time Collision Prevention for Electric Forklifts
In high-stakes industrial environments, safety is paramount. Our "Industrial Safety Edge AI" project addresses the critical challenge of preventing collisions involving electric forklifts, a leading cause of accidents in manufacturing and warehousing facilities. This innovative system leverages cutting-edge artificial intelligence deployed directly at the edge to provide real-time hazard detection and alert capabilities.
We partnered with KEAS, a major manufacturing plant, to implement and validate our solution. The system was meticulously integrated into their existing operations, focusing on minimizing disruption while maximizing safety impact. Our primary objective was to drastically reduce the incidence of forklift-related accidents, thereby protecting personnel, equipment, and operational continuity.
70%
Accident Reduction
Demonstrated decrease in forklift-related incidents at KEAS manufacturing plant within the first year of deployment, significantly improving workplace safety.
This successful deployment at KEAS showcased the system's effectiveness, delivering a tangible reduction in accidents and enhancing overall operational safety. The implementation proves that advanced AI can create safer workspaces, mitigate risks, and contribute to a more robust and efficient industrial landscape.
Problem Statement
Industrial workplaces, particularly those involving heavy machinery like forklifts, present significant safety challenges. According to OSHA, forklifts cause approximately 85 deaths and 34,900 serious injuries annually in the US. A significant portion of these incidents involve collisions with pedestrians, other vehicles, or static objects, leading to severe personnel injuries, equipment damage, operational downtime, and substantial financial losses. Traditional safety measures, such as mirrors, warning signs, and administrative controls, often prove insufficient in dynamic and complex industrial environments characterized by blind spots, varying lighting conditions, and human error. The need for an automated, real-time, and proactive collision prevention system is critical to establish a truly safe and efficient operational environment.
Solution Architecture
Our Industrial Safety Edge AI system is built upon a robust edge computing architecture designed for real-time performance and seamless integration. At its core, the system utilizes NVIDIA Jetson platforms, enabling powerful AI inference directly on the forklift.

  • NVIDIA Jetson Edge Deployment: We deploy the AI model on NVIDIA Jetson AGX Orin or Xavier modules, selected for their high AI inference capabilities, low power consumption, and industrial-grade reliability. These embedded systems provide the computational power required for real-time object detection at the source, eliminating latency associated with cloud processing.
  • Real-time Object Detection Pipeline: The core of the system is a highly optimized real-time object detection pipeline. Video feeds from multiple wide-angle cameras are ingested, processed by a YOLOv8-based model, which identifies and tracks pedestrians, other forklifts, and critical obstacles within predefined safety zones. The pipeline ensures frame-by-frame analysis with minimal latency.
  • Red Zone Alert System Mechanics: Detected objects are mapped into a 3D environmental model, enabling precise proximity calculations. When an object enters a configurable "red zone" (e.g., within 2 meters), the system activates multi-modal alerts: audible alarms (sirens, voice warnings), visual warnings (flashing lights, in-cabin displays), and haptic feedback to the operator.
  • Integration with Forklift Control Systems: For proactive collision prevention, the system integrates directly with the forklift's CAN bus (Controller Area Network) or digital I/O. This allows for automated speed reduction when an imminent collision is detected and, in critical situations, can initiate an emergency stop, significantly mitigating accident severity.
Technical Implementation
YOLOv8 Model Optimization for Edge Computing
The YOLOv8s (small) model was selected for its balance of accuracy and inference speed. We further optimized it for the Jetson platform through model pruning, reducing redundant layers and weights, and applying mixed-precision training (FP16) to minimize model size and computational load without significant accuracy degradation.
TensorRT Acceleration Techniques
NVIDIA TensorRT was extensively used to optimize the YOLOv8 model for maximum throughput on the Jetson GPUs. This involved creating an optimized inference engine by performing graph optimizations such as layer fusion, kernel auto-tuning, and INT8 quantization, which leverages the Jetson's specialized hardware accelerators.
Custom Dataset Creation and Training Process
A comprehensive dataset was curated comprising thousands of images and video frames from KEAS's facilities, annotated for various classes (pedestrians, forklifts, pallets, racks). Data augmentation techniques (e.g., random scaling, cropping, brightness adjustments) were applied to enhance model robustness. Training was conducted on NVIDIA V100 GPUs using PyTorch, with a learning rate scheduler and early stopping to prevent overfitting.
Multi-camera Setup and Calibration Procedures
The system employs a four-camera setup providing 360-degree coverage around the forklift. Each camera (e.g., industrial-grade GigE cameras) was factory calibrated for intrinsic parameters. Extrinsic calibration (camera-to-forklift and camera-to-camera) was performed using a chessboard pattern to enable accurate 3D position estimation of detected objects relative to the forklift's chassis.
Edge Computing Optimization
Model Quantization Techniques Used
Implemented Post-Training Quantization (PTQ) to convert floating-point model weights to INT8. This reduced model size by 75% and boosted inference speed by up to 2x on Tensor Cores, while maintaining a <2% drop in mAP.
Inference Speed Optimization Methods
Leveraged TensorRT for graph optimization, kernel fusion, and dynamic batching. Also employed NVIDIA's DeepStream SDK for efficient video stream processing, minimizing CPU overhead and maximizing GPU utilization.
Power Consumption Management Strategies
Configured Jetson devices to operate in optimized power modes (e.g., MAXN vs. 10W MODE) based on operational requirements. Dynamic Frequency Scaling (DFS) for CPU and GPU cores was implemented to adjust power draw according to real-time workload.
Real-time Processing Constraints and Solutions
Achieved end-to-end latency of <70ms (camera input to alert trigger) to ensure critical real-time responsiveness. This was met by balancing model complexity, quantization, TensorRT optimization, and efficient data handling.
Results & Performance Metrics
Beyond the initial 70% accident reduction, the system demonstrated robust technical performance during its deployment and validation phases:
95.2%
Detection Recall
For pedestrians and forklifts within 5 meters under varying conditions.
98.1%
Precision Rate
Minimizing false positives to avoid alert fatigue among operators.
30 FPS
Processing Throughput
Sustained across all camera streams per Jetson AGX Orin module.
65ms
End-to-End Latency
From frame capture to alert activation on average.
  • Reduced Accident Severity: Incidents that still occurred were predominantly minor, with a significant reduction in severe injuries and equipment damage.
  • Improved Operational Efficiency: While primarily a safety system, proactive alerts also led to better traffic flow in congested areas and reduced minor bumps and scrapes.
  • Enhanced Data Logging: The system logs all near-miss events, providing valuable data for continuous safety improvements and hotspot analysis.
Deployment Challenges
  • Varying Lighting Conditions: Industrial environments often have challenging and rapidly changing lighting, from bright sunlight to dim corners. This was mitigated by using high dynamic range (HDR) cameras and extensive data augmentation during training.
  • Dust and Debris: Cameras can accumulate dust and debris, obscuring vision. Regular cleaning protocols were established, and image processing filters were implemented to partially compensate for minor obstructions.
  • Forklift Vibrations: Constant vibrations from forklift operation can affect camera stability and image quality. Industrial-grade, vibration-resistant mounts were utilized, and software-based image stabilization techniques were integrated.
  • Integration with Legacy Systems: Interfacing with older forklift models that lack modern communication protocols like CAN bus required custom digital I/O modules and careful electrical engineering.
  • Operator Acceptance: Initial resistance from forklift operators to new technology and constant alerts required thorough training, clear communication of benefits, and fine-tuning alert sensitivity to prevent alert fatigue.
Technology Stack
1
Hardware
NVIDIA Jetson AGX Orin/Xavier, Industrial GigE Cameras (Basler/FLIR), CAN Bus Interfaces
2
Operating System
Ubuntu 20.04 LTS (JetPack 5.x)
3
AI Frameworks & Libraries
PyTorch (1.13+), YOLOv8, NVIDIA TensorRT (8.x), OpenCV (4.x), DeepStream SDK (6.x)
4
Programming Languages
Python (3.8+), C++
5
Communication Protocols
CAN Bus, MQTT (for data logging and remote monitoring)
6
Data Storage
SQLite (edge logging), AWS S3 (cloud archival)
Future Enhancements & Scalability
Our commitment to continuous improvement and broader impact drives our future development roadmap:
  • 3D Object Tracking: Implementing more advanced 3D object tracking algorithms to enhance prediction accuracy, especially in cluttered environments and for occluded objects.
  • Predictive Maintenance: Integrating sensor data from the forklift with AI to predict component failures and schedule proactive maintenance, reducing downtime.
  • Fleet Management Integration: Connecting individual forklift AI units to a central fleet management system for aggregated safety insights, real-time location tracking, and optimized traffic flow management across the entire facility.
  • Generative AI for Scenario Simulation: Utilizing generative AI to create synthetic data for training on rare accident scenarios, further improving model robustness without relying solely on real-world incident data.
  • Return on Investment (ROI) Analysis:
  • Direct Cost Savings: Significant reduction in costs associated with accidents, including medical expenses, equipment repair, insurance premiums, and legal fees.
  • Increased Productivity: Reduced downtime due to accidents and improved operational flow contribute to higher overall productivity.
  • Enhanced Employee Morale: A safer working environment leads to higher employee satisfaction, reduced turnover, and a more positive workplace culture.
  • Compliance & Reputation: Proactive safety measures ensure compliance with regulations and bolster the company's reputation as a responsible employer.
  • Scalability Model: The modular Jetson-based architecture allows for easy scalability across different forklift types and industrial facilities, with a clear cost-benefit analysis showing ROI typically within 12-18 months of full deployment, depending on fleet size and pre-existing accident rates.
Heritage Recognition AI: Instant Monument Identification for Cultural Tourism
Cultural tourism thrives on discovery and connection, yet a significant challenge persists: how can visitors instantly identify and understand the rich history embedded in the monuments around them? Our Heritage Recognition AI platform addresses this critical gap, transforming passive observation into immersive, educational experiences.
The Cultural Heritage Challenge
In an era of instant information, cultural tourists often find themselves grappling with outdated methods for monument identification. The experience of standing before a historical landmark, only to struggle with its name, origin, or significance, is all too common. This friction diminishes the potential for deeper engagement and a truly enriching cultural journey.
Traditional guidebooks can be cumbersome, and general internet searches often yield overwhelming or irrelevant results, especially for lesser-known sites. Many monuments, particularly in dense historical districts, bear strong resemblances to one another or lack clear, multilingual interpretive signage, leading to confusion and missed opportunities for learning.
This disconnect creates significant tourist experience gaps: visitors are left to passively observe rather than actively interact with the history before them. The spontaneity of discovery is frequently bogged down by the arduous process of information retrieval, hindering personalized learning and a profound appreciation for cultural heritage.
Project Overview
This project was initiated in partnership with the Ministry of Culture to revolutionize cultural tourism. Our primary goal is to enhance tourist engagement by providing an intuitive and instant method for identifying historical monuments. By bridging the gap between physical heritage sites and digital information, we aim to deepen visitors' appreciation and understanding of cultural history, transforming passive sightseeing into an interactive educational journey.
AI Solution Architecture
Our Heritage Recognition AI system leverages a sophisticated architecture designed for fast, accurate, and scalable monument identification. The core principle involves transforming visual information into machine-readable embeddings, enabling efficient comparison against a vast database of historical sites.
  • Deep Embedding Generation: At the heart of our system is a deep learning model responsible for generating unique, high-dimensional numerical representations (embeddings) for each monument image. These embeddings capture complex visual features, allowing for robust similarity comparisons.
  • FAISS Similarity Search Implementation: To enable real-time identification against a massive database of monuments, we utilize Facebook AI Similarity Search (FAISS). FAISS provides highly optimized algorithms for efficient similarity search and clustering of dense vectors, significantly speeding up the lookup process.
  • Mobile App Integration Architecture: The system is seamlessly integrated into a user-friendly mobile application. The app handles image capture, sends it to the backend for processing, and displays rich historical context and metadata received from the AI engine.
  • Real-time Image Processing Pipeline: An optimized pipeline ensures that captured images are pre-processed (e.g., resizing, normalization), fed to the embedding model, and the resulting vector is sent to the FAISS index for similarity search, all within milliseconds to provide an instant user experience.
Technical Implementation
CNN Architecture for Feature Extraction
We deployed a fine-tuned ResNet-50 convolutional neural network, pre-trained on ImageNet, as our feature extractor. The final classification layer was replaced with a global average pooling layer followed by a dense layer producing 512-dimensional embeddings. Further fine-tuning was conducted using triplet loss to optimize the embedding space for inter-class separation and intra-class compactness specific to monument recognition.
Embedding Space Optimization Techniques
Beyond triplet loss, we employed techniques such as ArcFace and CosFace to further enhance the discriminative power of our embeddings. These methods impose stricter angular margin penalties, resulting in more tightly clustered intra-class features and larger separation margins between different monument embeddings, crucial for distinguishing visually similar structures.
Database Indexing Strategies
For efficient FAISS indexing, we explored various strategies including IVF (Inverted File Index) for large-scale datasets, combined with PQ (Product Quantization) for memory reduction. The `IndexFlatL2` was used for smaller, highly accurate searches during development, transitioning to `IndexIVFPQ` for production-level performance with billions of vectors.
Mobile SDK Development Process
The mobile application was developed using native SDKs (Kotlin for Android, Swift for iOS), integrating custom APIs for image upload and data retrieval. Local caching mechanisms were implemented for frequently accessed information, and offline capabilities were planned by embedding a lightweight on-device model for basic recognition, complemented by the cloud-based deep learning backend.
Dataset Creation & Training
Monument Image Collection Methodology
Our dataset was compiled from diverse sources: high-resolution drone imagery, official Ministry of Culture archives, crowd-sourced contributions from historical societies, and partner museum collections. Each image was meticulously geo-tagged and cross-referenced with historical records.
Data Augmentation Techniques Used
To ensure model robustness, extensive data augmentation was applied. This included random rotations, shifts, zooms, shear transformations, brightness adjustments, contrast changes, and the simulation of various weather conditions (e.g., fog, rain) to generalize across diverse viewing environments.
Multi-angle and Lighting Condition Handling
The dataset explicitly includes images captured from multiple angles (e.g., front, side, aerial) and under varied lighting conditions (e.g., dawn, midday, dusk, overcast). This comprehensive approach minimizes the impact of viewpoint and illumination changes on recognition accuracy.
Historical Accuracy Validation Process
Every piece of historical information associated with an identified monument undergoes a rigorous validation process. This involves review by expert historians, cross-referencing with primary historical documents, and verification against established cultural heritage databases to ensure impeccable accuracy.
Performance Metrics
The Heritage Recognition AI platform consistently delivers high accuracy and speed, transforming the tourist experience with reliable and instant information.
97.8%
Recognition Accuracy
For trained monuments across diverse conditions.
150ms
Average Latency
From image capture to information display.
10M+
Indexed Monuments
Scalability demonstrated with a growing database.
94.5%
User Satisfaction
Based on post-pilot survey results.
Mobile Application Features
1
Instant Recognition
Point your camera at a monument and receive immediate identification and historical facts.
2
Rich Historical Context
Access detailed narratives, construction timelines, architectural styles, and cultural significance.
3
Multi-language Support
Information available in multiple languages to cater to a global tourist audience.
4
Interactive Maps & Itineraries
Navigate to nearby historical sites and discover curated walking tours.
5
Personalized Collections
Save identified monuments and create personal travel logs and galleries.
6
Augmented Reality Overlays
Visualize historical reconstructions or architectural details directly on the monument through the camera view.
Deployment & Scalability
The Heritage Recognition AI is designed for robust deployment and effortless scalability, ready to serve cultural tourists worldwide.
  • Cloud-Native Infrastructure: The backend is deployed on a cloud-native architecture (e.g., AWS, Azure, GCP), utilizing containerization (Docker, Kubernetes) for microservices, ensuring high availability and fault tolerance.
  • Geographic Distribution: Edge nodes and Content Delivery Networks (CDNs) are employed to minimize latency for users globally, ensuring rapid response times regardless of location.
  • Scalable Database: The monument embedding database scales dynamically to accommodate an ever-growing collection of heritage sites, using distributed database solutions.
  • API-First Design: All core functionalities are exposed via robust APIs, allowing for easy integration with third-party tourism platforms, smart city initiatives, and museum applications.
Technology Stack
1
Core AI Frameworks
PyTorch 2.0+, TensorFlow 2.x, FAISS 1.7.x
2
Backend Development
Python 3.9+, Flask/Django, FastAPI
3
Cloud Platform
AWS (EC2, S3, RDS, Lambda, SageMaker) / Azure / GCP
4
Mobile Development
Kotlin (Android SDK), Swift (iOS SDK)
5
Database
PostgreSQL 14, Redis 7
6
Containerization & Orchestration
Docker 20.10+, Kubernetes 1.25+
Cultural Impact & Future Enhancements
The Heritage Recognition AI platform extends beyond mere identification; it aims to foster a deeper connection between visitors and the world's cultural treasures, while offering avenues for continuous evolution.
  • Enhanced Educational Value: By providing instant, detailed, and accurate information, the platform significantly enhances the educational aspect of cultural tourism, making history accessible and engaging for all ages.
  • Preservation Awareness: Increased awareness and appreciation for monuments can contribute to greater public support for conservation and preservation efforts.
  • Economic Catalyst: Drives local tourism economies by encouraging exploration of lesser-known sites and extending visitor stay durations.
  • Crowd-Sourced Content Integration: Empowering users to contribute their own photos, historical anecdotes, and multilingual translations, further enriching the database and community engagement.
  • Personalized Learning Paths: Implementing AI-driven recommendations for monuments based on user preferences and previously identified sites.
  • Integration with Wearable Devices: Expanding the platform's reach to smart glasses and other wearables for an even more seamless and immersive hands-free experience.
  • 3D Reconstruction & Virtual Tours: Utilizing recognized monuments to generate or link to interactive 3D models and virtual tours, offering new ways to explore heritage from anywhere.



Metro Crowd Intelligence: AI-Powered Passenger Flow Management for Istanbul Transit
Istanbul, a vibrant metropolis straddling two continents, faces immense challenges in managing its burgeoning urban population and their daily transit needs. The Metro Crowd Intelligence System, developed for Istanbul's extensive metro network, is a groundbreaking AI solution designed to address these complexities. This project focuses on deploying advanced artificial intelligence at over 40 key metro stations to enhance passenger safety, optimize operational efficiency, and significantly improve the overall travel experience for millions of commuters.
Our system transforms raw data from existing infrastructure into actionable insights, providing real-time crowd dynamics, predicting potential bottlenecks, and enabling proactive management strategies. By leveraging cutting-edge computer vision and edge computing, we've created a robust, scalable, and highly accurate platform that redefines urban transit management in one of the world's busiest cities.
Urban Transit Challenge: Navigating the Dynamics of a Megacity
Istanbul's metro system, while crucial for urban mobility, grapples with several critical issues inherent to a rapidly growing megacity. During peak hours, stations become severely congested, leading to discomfort, extended wait times, and significant safety risks. The traditional methods of crowd management often rely on manual observation, which is inherently reactive and inefficient in fast-evolving environments.
These challenges directly impact operational efficiency, causing delays, increasing energy consumption, and straining personnel resources. More importantly, the sheer volume of passengers creates potential safety hazards, from minor incidents to severe crowd-related emergencies. The Metro Crowd Intelligence System was conceived to move beyond reactive measures, offering a predictive and real-time solution to maintain fluid passenger flow and ensure commuter safety.
AI Solution Architecture: Intelligent Eyes on the Ground
The core of the Metro Crowd Intelligence System lies in its distributed and intelligent architecture, designed for real-time processing and decision-making. At each of the 40+ deployed stations, a network of cameras continuously feeds visual data to compact, high-performance computing units. These units, primarily Raspberry Pi devices, are optimized for edge deployment, allowing for immediate, localized analysis without constant reliance on central servers. This minimizes latency and enhances system responsiveness.
The system integrates seamlessly with multi-camera setups, allowing for comprehensive coverage of platforms, escalators, and entrance/exit points. Real-time crowd counting algorithms, powered by deep learning models, process this visual information to accurately estimate crowd density and movement. All aggregated data and insights are then relayed to a central monitoring dashboard, providing transit authorities with a holistic, real-time overview of the entire metro network, enabling informed and timely interventions.
Technical Implementation: Powering Real-time Crowd Management
Computer Vision Models for Crowd Density Estimation
Utilizes state-of-the-art convolutional neural networks (CNNs) trained on diverse datasets of urban crowd scenes. Models are optimized for object detection (people counting) and density map regression, providing highly accurate estimations even in challenging lighting and occluded conditions.
Edge Computing Optimization for Raspberry Pi
Custom lightweight models and inference engines are deployed on Raspberry Pi 4 devices. Techniques like model quantization, pruning, and hardware acceleration (e.g., OpenCV with GPU support) are employed to ensure high-performance, low-latency processing directly at the station level.
Real-time Data Streaming Architecture
Employs efficient messaging queues (e.g., Kafka, MQTT) to stream processed data (crowd counts, density heatmaps, alert triggers) from edge devices to the central monitoring system. This ensures minimal data loss and continuous updates for operators.
Analytics and Reporting System
A powerful backend processes historical and real-time data to generate comprehensive analytics. This includes peak hour trends, station-specific congestion patterns, and performance metrics. Interactive dashboards provide visualizations for easy interpretation and operational planning.
Crowd Analysis Algorithms: Unpacking Movement Patterns
Beyond simple counting, our system employs sophisticated algorithms to understand and predict crowd behavior, offering a deeper level of intelligence for metro operators. These algorithms are the brain behind the proactive management capabilities, transforming raw visual data into predictive insights.
Density Estimation Techniques
Leveraging pixel-level density maps, our models provide granular information on crowd distribution. This allows for precise identification of overloaded areas, distinguishing between general congestion and potentially dangerous crush conditions. It's crucial for understanding where intervention is most needed.
Flow Pattern Recognition
The system tracks individual and group movements to identify common flow paths, bottlenecks, and anomalous behavior. By recognizing typical ingress and egress patterns, it can detect deviations that might signal a problem or an opportunity for efficiency improvements.
Bottleneck Detection
Utilizes spatial-temporal analysis to pinpoint areas where passenger flow is restricted. The algorithms analyze the accumulation rates and exit rates from specific zones to dynamically identify and even predict where bottlenecks are likely to form, allowing for preemptive action.
Predictive Analytics for Crowd Management
Incorporates historical data, real-time sensor inputs, and external factors (e.g., event schedules, weather forecasts) to predict crowd surges and potential congestion up to 30 minutes in advance. This foresight enables operators to deploy staff, adjust train schedules, or implement rerouting strategies proactively.
System Integration: A Unified Command Center
The Metro Crowd Intelligence System is not a standalone solution; it's designed for deep integration into existing metro operational frameworks, ensuring that its powerful insights translate directly into effective action. This seamless connectivity is vital for a comprehensive transit management strategy.
Metro Operations Center Integration
The central monitoring dashboard provides a unified view of crowd dynamics across the entire network, directly accessible by operations staff. It visualizes real-time data, historical trends, and predictive alerts on a single, intuitive interface, streamlining decision-making.
Real-time Alert Systems
Automated alerts are triggered when predefined crowd density thresholds are exceeded, or unusual flow patterns are detected. These alerts are sent to relevant personnel via multiple channels (e.g., dashboard notifications, SMS, email), ensuring rapid response.
Historical Data Analysis
All crowd data is archived and available for in-depth historical analysis. This allows for post-event review, long-term trend identification, and data-driven planning for future infrastructure upgrades, event management, and resource allocation.
Performance Monitoring
The system continuously monitors its own performance, including camera uptime, edge device health, data stream integrity, and algorithm accuracy. This ensures the reliability and continuous operation of the crowd intelligence platform.
Deployment Across 40+ Stations: Overcoming Complexities
Deploying an advanced AI system across a sprawling metropolitan metro network presents unique challenges, particularly when integrating with diverse existing infrastructure. Our approach prioritized robust planning and flexible solutions to ensure successful implementation across 40+ stations in Istanbul.
  • Installation Challenges and Solutions: Each station presented unique architectural layouts, lighting conditions, and power supply variations. Our team conducted detailed site surveys, customizing camera placements and edge device enclosures for optimal performance and minimal aesthetic impact. Solutions included weather-resistant casings for outdoor cameras and redundant power sources.
  • Network Infrastructure Requirements: A dedicated and secure network backbone was established to handle the high volume of real-time data streaming from edge devices to the central servers. This involved upgrading existing fiber optic lines and deploying secure VPNs to ensure data integrity and prevent unauthorized access.
  • Maintenance and Monitoring Protocols: A proactive maintenance schedule was implemented, incorporating remote diagnostics for edge devices and automated alerts for camera malfunctions. Local technical teams were trained for rapid on-site interventions, supported by a central monitoring team for system health oversight.
  • Scalability and Future-Proofing: The modular design of the system allows for easy expansion to additional stations and integration of new sensor types (e.g., thermal cameras, LiDAR) as the metro network evolves and new challenges emerge.
Results & Impact: Transforming Istanbul's Commute
The implementation of the Metro Crowd Intelligence System in Istanbul has yielded tangible benefits, fundamentally improving both the operational dynamics of the metro system and the daily experience of its passengers. Data collected during the pilot and initial rollout phases demonstrates significant advancements across key performance indicators.
18%
Congestion Reduction
During peak hours in monitored stations, leading to smoother boarding and alighting processes.
25%
Operational Efficiency Gains
Through optimized staff deployment and proactive management of potential incidents.
12%
Reduction in Minor Incidents
Such as falls and minor collisions, enhancing overall passenger safety across the network.
89%
Passenger Satisfaction
Reported by commuters feeling safer and experiencing reduced wait times.
7%
Energy Consumption Reduction
Due to more efficient train scheduling and platform management.
These improvements translate into a safer, more predictable, and more comfortable commuting experience for Istanbul's residents and visitors. The system's ability to predict and prevent issues rather than merely react to them marks a significant leap forward in urban transit management.
Technology Stack and Infrastructure
The robustness and scalability of the Metro Crowd Intelligence System are built upon a meticulously selected technology stack, leveraging industry-leading tools and platforms optimized for real-time AI inference and big data processing in a distributed environment.
1
Edge AI Frameworks
TensorFlow Lite, PyTorch Mobile, OpenVINO for optimized model inference on Raspberry Pi devices.
2
Computer Vision Libraries
OpenCV, Dlib for image processing, camera calibration, and low-level vision tasks.
3
Data Streaming & Messaging
Apache Kafka for high-throughput, fault-tolerant data pipelines from edge to cloud. MQTT for lightweight communication between edge devices and local gateways.
4
Backend & API Development
Python with FastAPI for high-performance APIs, Node.js for real-time WebSocket communication to the dashboard.
5
Cloud Infrastructure
Microsoft Azure (Azure IoT Edge, Azure Kubernetes Service, Azure Stream Analytics, Azure Cosmos DB) for scalable data storage, processing, and management.
6
Database
PostgreSQL for structured operational data, InfluxDB for time-series sensor data, Redis for caching and real-time data lookups.
7
Monitoring & Visualization
Grafana for custom dashboard creation and real-time data visualization, Prometheus for system metrics collection.
Scalability and Future Smart City Applications
The Metro Crowd Intelligence System is inherently designed for scalability, capable of expanding its footprint beyond the initial 40+ stations and integrating with a broader smart city ecosystem. Its modular architecture ensures that new features and increased data volumes can be accommodated seamlessly.
  • Horizontal Scalability: The distributed edge computing model allows for easy addition of more Raspberry Pi units and cameras to new stations or existing ones for increased coverage without overburdening central servers.
  • Smart City Integration: The platform's API-first design facilitates integration with other urban management systems, such as traffic control, public safety, and event management platforms. For example, metro crowd data can inform smart traffic light adjustments or trigger police presence in specific areas.
  • Predictive Urban Planning: Aggregated, anonymized crowd movement data can be invaluable for urban planners, informing decisions on infrastructure development, public space design, and emergency evacuation routes.
  • Enhanced Emergency Response: In crisis scenarios, the system can provide real-time information on crowd distribution and movement, crucial for guiding emergency services and coordinating evacuations efficiently.
  • Dynamic Resource Allocation: Beyond metro operations, the intelligence derived can inform dynamic allocation of resources like public buses, sanitation services, or even commercial advertising based on real-time population density.



Smart City Solutions
Metro Crowd Monitoring
Engineered real-time passenger density monitoring using crowd-counting AI on Raspberry Pi at 40+ Istanbul metro stations, enhancing commuter experience and reducing congestion.
  • Real-time density tracking
  • Congestion reduction algorithms
  • Enhanced passenger flow
N-GPT RAG System
Created N-GPT, an LLM-powered RAG system using LLaMA for chatting with complex land-related PDFs, automatically parsing legal and geographic content for natural language interaction.
  • Document parsing automation
  • Natural language queries
  • Legal content analysis
Impact Metrics & Achievements
70%
Accident Reduction
Achieved through edge AI safety solution for industrial forklifts
90%
Processing Time Saved
V-LLM complaint system auto-detection efficiency improvement
40+
Metro Stations
Deployed crowd-counting AI across Istanbul's transportation network
30%
Customer Engagement
Increase through in-store analytics and movement pattern tracking
Technology Stack & Tools
Core Frameworks
PyTorch, TensorFlow, YOLOv5/v8/v11, OpenCV, and Hugging Face Transformers for advanced AI model development.
Edge Computing
NVIDIA Jetson, DeepStream 6.0, TensorRT optimization for real-time processing on embedded systems.
AI Systems
LangChain, FAISS, SAM/SAMv2, LLaMA, and RAG systems for intelligent document processing and retrieval.
Deployment
Docker, Streamlit, GeoPandas for scalable AI application deployment and geographic data processing.
Education & Professional Development
01
Master's Degree
Medipol University - Electric Electronics Engineering and Cyber Systems (2019-2021)
02
Bachelor's Degree
COMSATS University - Computer Science, AI and Machine Learning (2015-2019)
03
Research Experience
Medipol University - Machine Learning Researcher focusing on wide area surveillance (2019-2021)
04
Industry Consulting
Ant Media - Computer Vision Consultant specializing in image segmentation and 3D reconstruction
Let's Connect & Collaborate
Ready for New Opportunities
Open to Computer Vision Engineering and Industrial AI opportunities. Passionate about transforming AI research into real-world solutions that make cities smarter and workplaces safer.
Specializing in edge AI deployment, smart city solutions, and industrial automation with proven track record of delivering impactful results.
Key Certifications
  • Convolutional Neural Networks in TensorFlow
  • Disaster Risk Monitoring Using Satellite Imagery
  • Prompt Engineering for Developers
  • Building Video AI Applications at the Edge
  • Multimodal RAG: Chat with Videos!
Industrial Automation & Computer Vision Projects
Volumetric Calculation via Drone Imagery
Implemented an AI system to perform accurate volumetric calculations on construction sites using drone-captured imagery. This enhanced inventory management and progress tracking.
Automated Pick-and-Place Robotics
Developed computer vision algorithms for precise object detection and localization, enabling robotic arms to perform automated pick-and-place operations in manufacturing.
Automated Quality Control
Designed and deployed an AI solution for real-time defect detection and quality assurance in industrial production lines, significantly reducing manual inspection errors.
Made with