GenAI Forecasting with AWS & TimeLLM

A scalable retail forecasting system using GenAI (TimeLLM), AWS Glue, SageMaker, and Streamlit to deliver real-time, explainable predictions for smarter inventory and supply chain planning.

Comment

Inspiration

Retail businesses often struggle with demand fluctuations, overstocking, and understocking, leading to significant revenue loss. Inspired by this real-world challenge, I set out to build a modern, scalable, and AI-powered forecasting solution that bridges the gap between data engineering and cutting-edge machine learning.

What It Does

This project delivers an end-to-end forecasting pipeline for retail supply chains, leveraging both traditional ML and GenAI. It processes raw transactional and store-level data, transforms it into model-ready features, and generates weekly sales forecasts using both XGBoost and TimeLLM (a GPT2-based time series model). The results are visualized in an accessible Streamlit dashboard with explainability and voice interaction.

How I Built It

The solution follows a modular data engineering architecture powered by AWS:

  • Data Lake: Raw CSVs stored in Amazon S3, cataloged with AWS Glue Crawlers
  • ETL Pipeline: Built in AWS Glue Studio to clean, join, and transform data into Parquet format
  • Feature Store: Partitioned datasets used to create model-ready inputs using Athena
  • Modeling:
    • Trained XGBoost models grouped by store_type
    • Integrated TimeLLM model in GPT2 mode for GenAI-based time series forecasting
  • Inference Layer: Batched and real-time predictions via FastAPI + Lambda
  • Visualization: Forecasts, comparisons, and explainability surfaced via Streamlit with voice support
  • Feedback & Security: Prediction logging in RDS, with IAM, CloudWatch, and Macie for monitoring and compliance

Challenges

  • Integrating a custom GenAI time series model (TimeLLM) within AWS SageMaker
  • Handling large-scale joins and partitioning in Glue for performance
  • Balancing interpretability with accuracy across ML and GenAI approaches
  • Ensuring real-time responsiveness and accessibility in the frontend

What I Learned

  • Advanced data pipeline design using AWS services
  • Deploying and integrating LLMs for time series use cases
  • Building secure, explainable, and production-grade ML systems
  • Comparing traditional ML with GenAI to evaluate trade-offs in accuracy, transparency, and complexity

What's Next

  • Fine-tune the TimeLLM model on retail-specific sequences
  • Add real-time alerting and budget planning features
  • Open source the architecture for educational and enterprise use

Built With

Share this project:

Updates