Project Advisor (Proje Danışmanı):
Araş. Gör. Dr. Gözde KUŞ
Project Lead / Principal Investigator (Proje Yürütücüsü):
Meriç Özcan
Project Team Members (Proje Ekip Üyeleri):
- Buket Gürlek — Front-end Development (UI / UX, Web Interface)
- Mirgavam Gavami — Model Backend Development (In Progress)
This repository presents a TÜBİTAK 2209-A–level, end-to-end artificial intelligence–based decision support system developed to predict, optimize, and manage operational risks in secure cargo delivery, last-mile logistics, and product return processes.
The project focuses on replacing reactive operational approaches with a proactive, risk-aware decision framework by integrating machine learning–based risk prediction, probability optimization, and interpretable model outputs.
The resulting system is academically rigorous, methodologically sound, and directly applicable to real-world operational environments.
The entire lifecycle of the project—from data preparation to model validation and final presentation—was executed in compliance with Agile project management principles and academic research ethics.
Artificial Intelligence · Risk Prediction · Secure Cargo Delivery · Last-Mile Logistics · Product Return Risk · Explainable Machine Learning · Decision Support Systems · Agile Project Management
The primary objective of this project is to develop a web-oriented, AI-driven decision support infrastructure capable of:
- Predicting product return risk at the individual order level
- Supporting secure cargo and logistics optimization decisions
- Producing interpretable and actionable risk outputs
- Enabling process-level proactive risk management
- Maintaining academic validity and reproducibility
The system is designed to assist operational stakeholders before shipment decisions are finalized, thereby reducing cost, inefficiency, and uncertainty.
The system follows a modular and scalable architecture, ensuring robustness and extensibility.
- Customer, order, category, logistics, and process-related variables
- Leakage-aware preprocessing and feature engineering
- Explicit handling of class imbalance
- Supervised machine learning–based risk prediction
- Risk score generation and categorical risk classification
- Comparative evaluation across alternative algorithms
- Emphasis on generalizability over metric maximization
- Probability threshold optimization aligned with operational costs
- Iterative feedback loop for underperforming model paths
- Decision-oriented risk calibration
- Human-interpretable outputs:
- Risk Score
- Risk Class
- Dominant Contributing Variables
- Operational Interpretation
- Academic reporting and colloquium-ready visualization outputs
The project was managed using Agile methodology, ensuring continuous validation and transparent progress tracking.
flowchart TD
A[System Initialization]
B[Data Collection and Evaluation]
C[Database Preparation]
D[Algorithm Development and Validation]
E[Probability Optimization]
F[Model Revision Loop - Feature and Model Update]
G[Web and Interface Integration]
H[System Testing and Optimization]
I[Final Delivery and User Presentation]
A --> B
B --> C
C --> D
D -->|Successful| E
D -->|Unsuccessful| F
F --> D
E --> G
G --> H
H --> I
| File | Description | Author |
|---|---|---|
| 2209-A General Work Plan 15.08.25.pdf | TÜBİTAK-style general project plan | Meriç Özcan |
| August and September 2025 Work Plan.pdf | Agile sprint planning and timeline | Meriç Özcan |
| Return_Risk_Analysis.ipynb | Demo machine learning pipeline | Meriç Özcan |
| Risk Analysis in Maritime Transportation_Literature Review.pdf | Academic literature foundation | Meriç Özcan |
| Analysis of Maritime Accidents Using Bayesian Networks.html | Probabilistic risk modeling reference | Meriç Özcan |
| Last Mile Logistics Optimization.html | Logistics optimization framework | Meriç Özcan |
| An Overview of Management in Maritime Transportation.html | Managerial and operational context | Meriç Özcan |
| Colloquium Presentation.pdf | Academic presentation material | Meriç Özcan |
| Colloquium Written Statement.pdf | Formal written academic contribution | Meriç Özcan |
| index.html | Web interface entry point (project landing page) | Buket Gürlek |
| script.js | Front-end logic and interactivity | Buket Gürlek |
| style.css | Front-end styling and layout | Buket Gürlek |
| Backend (in progress) | Model backend development and integration | Mirgavam Gavami |
Model evaluation was conducted with a methodology-first perspective, where predictive performance is assessed alongside:
- F1-Score and class stability
- Sensitivity to data leakage
- Robustness across validation splits
- Interpretability and decision relevance
The project explicitly avoids black-box optimization, prioritizing explainable and auditable modeling choices.
Each prediction generated by the system includes:
- Continuous risk score
- Discrete risk class (Low / Medium / High)
- Key contributing features
- Operationally meaningful interpretation
This design ensures that model outputs support human decision-making rather than replace it, aligning with real operational and managerial needs.
The finalized system successfully delivers:
- A validated and leakage-aware ML pipeline
- Interpretable, decision-ready risk outputs
- Agile-compliant development documentation
- Reproducible academic deliverables
- Presentation-ready decision support artifacts
The project was formally concluded upon achieving:
- Fully integrated data, model, and decision layers
- Verified model robustness and interpretability
- Complete academic documentation
- Agile sprint closure with defined deliverables
This repository demonstrates a holistic approach to AI-driven risk management, combining:
- Academic discipline
- Engineering best practices
- Agile methodology
- Decision-oriented analytics
It reflects not only what was built, but how a serious, production-ready AI system should be designed and validated.
This work is intended for academic, educational, and research purposes.
Commercial usage requires additional validation and domain-specific adaptation.