About This Project

A journey through research, engineering, and innovation in time series anomaly detection

ByHarshvardhan Santosh MagarGitHub Repository

Project Timeline

Week 1

Problem Research & Literature Review

Investigated time series anomaly detection challenges across e-commerce, IoT, and security domains. Reviewed 20+ papers on autoencoders, attention mechanisms, and threshold learning.

Week 2

Architecture Design & Synthetic Data

Designed hybrid LSTM-Transformer autoencoder with adaptive threshold network. Generated 100K-timestep synthetic dataset with 3 realistic anomaly types and comprehensive baselines.

Week 3

Training, Evaluation & Ablation Studies

Trained models on GPU, ran comprehensive benchmarks (Precision/Recall/F1/AUC-ROC), proved 5.6% improvement from adaptive thresholds via ablation study.

Week 4

Web Platform & Research Paper

Built professional web app with interactive demo, real-time inference, and benchmarks dashboard. Wrote arXiv-ready research paper with methodology and results.

Technology Stack

🧠 Machine Learning

  • • PyTorch - Deep learning framework
  • • LSTM & Transformer layers - Neural architectures
  • • Scikit-learn - Baseline models & metrics
  • • NumPy/Pandas - Data processing

🌐 Web Development

  • • Next.js 16 - React framework
  • • TensorFlow.js - Browser inference
  • • Recharts - Data visualizations
  • • Tailwind CSS - Styling

☁️ Infrastructure

  • • Google Colab - Model training
  • • Vercel - Web deployment
  • • GitHub - Version control
  • • jsPDF - PDF export

🔍 Research

  • • LaTeX - Paper formatting
  • • KaTeX - Equation rendering
  • • Statistical analysis - Ablation studies
  • • Matplotlib - Publication figures

Key Technical Insights

1. Hybrid Architectures Balance Multiple Objectives

Combining LSTM (local patterns) with Transformers (global context) outperforms either alone. This teaches the importance of architectural diversity—no single model dominates all problem dimensions.

2. Context-Aware Learning Beats Fixed Heuristics

Learning per-timestep thresholds instead of using fixed values improved F1 by 5.6%. This demonstrates that domain-specific learning (even simple neural networks) can outperform hand-crafted rules in production systems.

3. Ablation Studies Are Non-Negotiable

Rigorously removing each component (LSTM, Transformer, adaptive thresholds) proved their individual contributions. This scientific rigor is essential for publication and demonstrates true understanding.

4. Reproducibility Requires Discipline

Fixed random seeds, documented hyperparameters, and multiple runs with error bars ensure results are trustworthy. This is why production systems demand careful engineering, not just model architecture.

Try It Yourself

All code is open source and fully reproducible. Train the model yourself in Google Colab—no GPU required (though recommended for speed).

Ready to explore the model?