Portfolio
AI Systems and Proof-of-Concept Projects
Built with Advanced Research and Engineering
Humanization LLM-Based Characters by Background Story Injection
A mobile communication app built on Large Language Models (LLMs). It uses character-specific context and behavioral frameworks, defined by a professional, to allow users to engage with human-like AI characters ("AI Peers") during the orientations they are currently in. The infrastructure relies on Microsoft Foundry and utilizes OpenAI APIs supported by a vectoral database.
Intensely Layered Data Visualization for a Digital Twin of Networking Data Center
A cloud-enabled architecture modeling a data center's hierarchical structure (floors, racks, rows) host 5G networks. It uses event-driven and streaming mechanisms to integrate real-time KPIs (latency, temperature) for proactive management. The interface utilizes node-based libraries to reduce visual complexity of data and relies on Neo4j for database management.
Synthetic Visual Data Generation For Drone-Based Network Tower Inspection & Digital Twin
An innovative solution combining computer vision and industrial automation. It processes drone video using advanced image processing models to create a high-fidelity visual digital twin with over 95% equipment detection accuracy. To prove the concept, synthetic data generated from Unreal Engine 5 with 3D tower models in the absence of drone footage. Then, the synthetic video is processed within RealityScan to generate a digital twin. Object detection is applied to identify tower equipment from very limited samples.
Visual Data Collection From Remote Participants
A React Native mobile framework for iOS/Android that uses gamification to drive large-scale, ethical video data collection for sign language research. A custom sign data collection app is built and distributed among the deaf community and volunteer participants record their hand gestures through phone cameras by the in-app instructions. Visual feature extraction is applied on the collected data by utilizing the MediaPipe framework. The selected features are processed by secure, scalable ML pipelines including in the cloud for gesture recognition and model training.
Real-Time Sign Language Translation In Uncontrolled Environments
A React Native hybrid app providing real-time translation via a low-latency architecture. It uses WebSocket communication to stream video frames to high-performance FastAPI backends for rapid ML-based gesture recognition.
Visual Data Collection For Eye-Tracking From Remote Participants
A custom eye data collection app is built and distributed among the ALS/SM community and volunteer participants record eye gaze through commodity webcams by the in-app instructions. The collected data trained an accurate model to map eye gaze from standard cameras to screen coordinates for real-time cursor control via a lightweight desktop application.
Audial Data Collection For Dysarthric Speech Recognition From Remote Participants
A custom speech data collection app is built and distributed among the dysarthric community and volunteer participants record their speech through commodity microphones by the in-app instructions. Specialized speech recognition is applied on the collected data for dysarthric Turkish speech. It utilizes an RMS-based silence detection and waveform segmentation data pipeline to train a model achieving a Word Error Rate (WER) of under 25%.
Blockchain Airdrop Fraud Detection
A high-performance fraud detection system that curates specialized datasets from restricted blockchain data. It utilizes sophisticated machine learning models to identify subtle fraudulent activity patterns in Web3 and fintech ecosystems. The detection algorithm relies on Graph Neural Networks (GNNs).