AI Portfolio⚡

TL;DR

My portfolio of applied AI and machine learning projects demonstrating various techniques across time series forecasting, natural language processing, traditional machine learning and data visualisation tasks, currently including:

Time Series

• XGBoost: Gradient boosting implementation for time series prediction
• Prophet: Facebook's forecasting tool for time series with seasonality

Capabilities

• Multi-horizon forecasting (e.g., 1-day, 7-day, 30-day ahead)
• Comprehensive metrics: MAE, RMSE, MAPE, SMAPE, R², MASE
• Automated train/validation/test splitting with temporal ordering
• Feature engineering (lag features, date components)
• Model comparison with detailed visualisations
• Persistent model storage and results tracking

Product Reviews Classification

• Machine learning based models: logistic regression, naive bayes
• Deep learning based models: feed forward neural network, CNN, RNN (LSTM), GRU,Bi-LSTM with attention
• Pre-trained language models:
  • BERT, RobERTa, DistilBERT (base & fine-tuned)

Data Visualisation

Available at edin-vis.streamlit.app

General Features

Since my main goal here is learning and gaining experience, I have tried to stick to the below principles:

• Modular Architecture: Clean separation between different ML domains with reusable pipeline components
• Production-Ready Code: Well-structured, documented, and maintainable implementations
• Comprehensive Evaluation: Multiple metrics, visualisations, and comparative analysis
• Configuration-Driven: Easy experimentation through YAML configuration files
• Extensible Design: Framework supports adding new models and project types
• Modern Libraries: Uses well-known python packages such as wandb,mlflow, pytorch,lightning, transformers, huggingface-hub, accelerate,unsloth, dspy, langchain, weaviate

Project Structure [ongoing]

AI-Portfolio/
├── TimeSeries/              # Time series forecasting
│   ├── configs/
│   │   └── config.yaml      # Model and pipeline configuration
│   ├── data/
│   │   ├── raw/             # Original time series data
│   │   └── processed/       # Processed datasets
│   ├── models/              # Model implementations
│   │   ├── statistical/     # Prophet, ARIMA, ETS, Theta
│   │   ├── ml/              # XGBoost, LightGBM, Random Forest
│   │   └── neural/          # LSTM, GRU, N-BEATS, Transformer
│   ├── src/                 # Core pipeline code
│   │   ├── data_loader.py   # Data loading and preprocessing
│   │   ├── pipeline.py      # Main pipeline orchestration
│   │   ├── evaluate.py      # Evaluation and metrics
│   │   └── utils.py         # Utility functions
│   ├── results/             # Output storage
│   │   ├── models/          # Saved models
│   │   ├── figures/         # Visualisations
│   │   └── metrics/         # Performance metrics
│   ├── main.py              # Entry point
│   └── requirements.txt     # Dependencies
├── NLP/                     # Natural Language Processing tasks
│   ├── configs/
│   │   └── config.yaml      # Model and pipeline configuration
│   ├── data/
│   │   ├── books/   
│   │   ├── dvd/
│   │   ├── electronics/
│   │   └── kitchen_&_housewares/   
│   ├── models/              # Model implementations
│   │   ├── Deep/            # FFNN, CNN, RNN, Transformers
│   │   ├── LMs/             # BERT, RoBERTa, DistilBERT
│   │   └── ml/              # Logistic Regression, Naive Bayes
│   ├── src/                 # Core pipeline code
│   │   ├── data_loader.py   # Data loading and preprocessing
│   │   ├── pipeline.py      # Main pipeline orchestration
│   │   ├── evaluate.py      # Evaluation and metrics
│   │   └── utils.py         # Utility functions
│   ├── results/             # Output storage
│   │   ├── models/          # Saved models
│   │   ├── figures/         # Visualisations
│   │   └── metrics/         # Performance metrics
└── MachineLearning/         # Traditional ML tasks (PLANNED)

Setup

1. Navigate to the desired directory (TimeSeries, NLP/Classification):

cd DIRECTORY

1. Install dependencies:

uv python install 3.12
uv venv --python 3.12
source .venv/bin/activate   # optional

uv pip install -r requirements.txt

Usage

Run the pipeline for each sub-project with default settings:

python main.py

Configuration

Edit **/configs/config.yaml in each sub-project. E.g. for the Time Series project you can customise:

• Data Settings: File paths, timestamp column, target variables
• Model Parameters: Hyperparameters for XGBoost and Prophet
• Forecast Horizons: Which time steps ahead to predict (e.g., [1, 7, 30])
• Evaluation Metrics: Which metrics to compute

Example configuration for XGBoost:

models:
  ml:
    xgboost:
      n_estimators: 100
      max_depth: 6
      learning_rate: 0.1

Experiment Configuration

Define experiments in **/main.py e.g.:

EXPERIMENTS = [
    {
        "model": "xgboost",
    },
    {
        "model": "prophet",
    },
]

Results

After running the pipeline, results are saved to:

• Models: **/results/models/
• Metrics: **/results/metrics/results_summary.json
• Figures: **/results/figures/

Latest Results

Below are the latest model comparison results from the time series forecasting pipeline comparing XGBoost and Prophet across multiple forecast horizons.

Model Comparison by Horizon

Comprehensive comparison of model performance across different forecast horizons (1-day, 7-day, 30-day ahead).

Single Horizon Performance (H=1)

Detailed performance metrics for 1-step ahead forecasting.

Test Set Predictions

Visual comparison of actual vs predicted values on the test set across all forecast horizons.

Pending⏱️

Time Series

• Additional statistical models: ARIMA, ETS, Theta
• Machine learning models: LightGBM, Random Forest, Gradient Boosting
• Deep learning models: xLSTM, GRU, N-BEATS, Transformer, TCN
• Advanced ensemble methods
• Hyperparameter optimisation

NLP

• Multi-modal agentic VQA with open-source VLMs
• More open source [thinking] models for classification (nemotron, mistral, qwen, etc. for zero-shot, few-shot, and chain-of-thought prompting)