AI-Driven Inventory Optimization & Demand Forecasting

• Synthetic Data Engineering: Generated a realistic FMCG dataset simulating daily raw material consumption, incorporating seasonality, trend factors, and random noise to mimic real-world factory volatility.
• Time-Series Feature Engineering: Transformed raw consumption data into a supervised learning problem by engineering features such as Lag Values (t-1, t-7), Rolling Averages, and volatility indices. This allowed the model to capture complex temporal patterns that traditional linear methods miss.
• Predictive Modeling (XGBoost): Implemented an Extreme Gradient Boosting (XGBoost) regressor to forecast future demand. The model significantly outperformed traditional Moving Average baselines, minimizing the Root Mean Square Error (RMSE).
• Dynamic Inventory Optimization: Developed a custom algorithm to calculate Dynamic Safety Stocks. Instead of a fixed buffer, the system calculates safety stock based on the forecast error of the model. This ensures higher buffers during volatile periods and leaner stocks during stable periods.

• Languages: Python 3.x
• Machine Learning: XGBoost, Scikit-learn
• Techniques: Time Series Forecasting, Regression, Feature Engineering, Dynamic Safety Stock Calculation, Hyperparameter Tuning
• Data Science: Pandas, NumPy
• Visualization: Matplotlib, Seaborn (for Forecast vs. Actual & Cost Analysis plots)

Classification of Iron Ionization States Using Ensemble Learning on NIST Spectral Data

• Automated Data Pipeline: Built a dynamic web scraper using Python to extract raw spectral data from the NIST server. Implemented complex Regex (Regular Expression) cleaning to handle measurement uncertainties and noise in raw wavelength/intensity data.
• Physics-Informed Feature Engineering: Beyond standard features, I derived a new physical attribute, "Energy Difference (ΔE)," representing the gap between upper and lower energy levels, which significantly improved model separability.
• Comparative Model Analysis: Evaluated three distinct algorithms: Naive Bayes (Baseline), Support Vector Machines (Kernel-based), and Random Forest (Ensemble). Validated results using 10-Fold Cross-Validation to ensure generalization.
• Results: The Random Forest model outperformed others, achieving a 99.16% Accuracy and 99.15% F1-Score, successfully distinguishing complex spectral patterns where the baseline model struggled.

• Languages: Python 3.x
• Machine Learning: Scikit-learn (Sklearn)
• Techniques: Random Forest, SVM (RBF Kernel), Naive Bayes, K-Fold Cross Validation, Web Scraping, Feature Engineering
• Data Science: Pandas, NumPy
• Visualization: Matplotlib, Seaborn, Heatmaps

Remaining Useful Life Prediction of Turbofan Engines Using LSTM Networks with Piecewise Linear Target Labeling

• Deep LSTM Architecture: Engineered a stacked LSTM network (128 & 64 units) with Dropout regularization to model long-term temporal dependencies in multivariate sensor data.
• Piecewise Linear Strategy: Applied a target clipping technique (RUL limit at 125 cycles) to stabilize training during the engine's "healthy" phase, significantly improving model convergence.
• Production-Grade Engineering: Transformed raw research notebooks into a modular, object-oriented codebase (src, models, api).
• Microservice Deployment: Developed a high-performance REST API using FastAPI, enabling external systems to send sensor data and receive RUL predictions in milliseconds.
• Results: Achieved a Root Mean Squared Error (RMSE) of 13.87 on the FD001 test set, outperforming standard baselines with a robust and deployable architecture.

• Core: Python 3.10+
• Deep Learning: TensorFlow, Keras, LSTM
• Deployment & API: FastAPI, Uvicorn, REST Architecture
• Data Engineering: Pandas, NumPy, Scikit-learn (MinMax Scaling, Windowing)
• Tools: Git, Modular Design, Swagger UI

DataDiet Daily AI Newsletter

KUKA Robot Dynamics Analysis

6-DOF Robot Arm Forward Kinematics Simulator

 Real-Time Sensor Monitoring & ML-Based Fault Detection

Physics-Informed ML: Airfoil Noise Prediction

• Physics-Aware Pipeline: Developed a custom data loader to validate aerodynamic constraints (e.g., boundary layer thickness) before training.
• Interactive Simulation: Deployed a Streamlit dashboard, allowing users to visualize noise profiles under different flight conditions in real-time.
• Clean Code: Structured the project as a modular Python package suitable for scalability.

End-to-End Autonomous Driving: Lane Detection & Steering Control

High-Altitude Model Rocket Design & Simulation