Minotask Core contains the experimental and production-ready Jupyter notebooks that power the IntriBot, an AI-assisted workflow for preparing Intrastat trade declarations. The notebooks orchestrate Google Generative AI models together with PDF parsing utilities to transform raw invoice documents into a structured commodity summary that can be uploaded to national Intrastat portals.
Compiling Intrastat reports traditionally requires finance or logistics teams to read through stacks of invoices, identify the goods involved, and manually assign Combined Nomenclature (CN) codes, origin countries, weights, and values. This manual process is slow, error-prone, and difficult to scale when transaction volumes increase or when product catalogs frequently change.
The project aims to automate as much of the Intrastat preparation workflow as possible by:
- Building a curated knowledge base of known products and their Intrastat metadata.
- Using large language models to classify invoice line items, infer missing details (such as commodity codes or weights), and cross-check for consistency.
- Producing a clean, reviewable export that finance teams can validate and submit.
Minotask_intriBot.ipynb– Interactive prototype notebook that uploads invoice PDFs, extracts key data, and generates a preliminary Intrastat summary using prompt-engineered LLM calls.Minotask_intriBot_prod_v1.ipynb– Production-focused notebook with modular cells for building the product knowledge base, configuring the AI prompts, and running the end-to-end processing pipeline against live data.
Both notebooks are designed to run in Google Colab. Follow the "How to Run It" section at the end of each notebook for step-by-step execution instructions, including API key configuration and expected inputs.