How petr-ngn/deepnote-hackathon works
Project type and purpose
• This is a full-stack Python application (a single-page Streamlit UI front-end plus a backend orchestrator) whose purpose is to extract financial data from scanned Czech company PDFs, enrich it with web-scraped company info, run LLM-powered financial analysis, and present results in an interactive dashboard.
• Key features: AWS Textract OCR processing, LLM integration via Amazon Bedrock and Anthropic Claude, web scraping with the Tavily API, AWS S3 storage, and a Streamlit-based UI.
File structure
• Top level:
• app.py — entrypoint which initializes Streamlit UI and kicks off processing
• Dockerfile, .dockerignore, pyproject.toml — project build & packaging
• config/ — YAML configs for LLM, OCR, scraper, UI
• src/ — core modules: aws.py, ocr.py, llm.py, scraper.py, ui.py, utils.py
• .github/workflows/pylint.yml — CI linting
• .pdf_examples/ — sample PDFs for demo
• Patterns:
• “src” implements a modular service-oriented architecture: each Python module encapsulates one external integration
• app.py and src/ui.py provide the UI layer; src/* share helper functions in utils.py
• config YAMLs parameterize external services, allowing easy swap or reconfiguration
Extract insights from README
• External services: AWS Textract, S3, Amazon Bedrock LLMs, Anthropic Claude, Tavily search API.
• Streamlit front-end drives file upload, displays progress spinners, and shows summary tables.
Essential architectural elements
• Service endpoint, data flow, and the orchestration sequence.
• Configuration injection (which drives credentials, adapter IDs, UI settings).
• Docker/Deepnote deployment boundary around the app & config files.