Income Tax Memo Automation
Budget: $20 – $25 USD
I’m kicking off the first block of a long-term income-tax automation program and need a Python engineer who can stand up a Prefect-based pipeline that automatically drafts tax memos. In this phase we are zeroing in on the drafting step itself; risk checking and further enrichment will follow in later blocks.
Here’s the workflow you will build:
• Inputs arrive through automated ingestion (think API push or scheduled fetch) as already structured tax fact patterns plus a skeletal memo outline.
• A Prefect flow then orchestrates a chain of LLM modules, RAG retrieval, and enrichment tasks.
• The flow returns a JSON-wrapped memo draft that a human reviewer can open immediately in our internal tools.
You’ll be free to choose the most appropriate vector store, embedding model, and validation strategy so long as everything runs on Python 3.9+ and is fully test-covered. Clean code, tight unit tests, and concise documentation are critical because future engineers will extend this work into risk identification, context-wrapper expansion, and prompt refinement.
Deliverables
- A reproducible Prefect flow (v2) in a Git repo
- Supporting modules for LLM calls, RAG retrieval, and data validation
- Unit tests that cover the core logic (pytest preferred)
- A short README explaining setup, configuration, and how to trigger the flow locally and in a Prefect deployment
Acceptance criteria
- End-to-end run produces a JSON file that mirrors the provided schema and contains a plausibly-structured memo body
- All tests pass with `pytest -q`
- Code quality meets black/flake8 standards and is type-hinted
If you thrive on architecting clean, production-ready data pipelines and know Prefect, LLM integration, and RAG inside out, I’d love to see how you’d approach this.
Here’s the workflow you will build:
• Inputs arrive through automated ingestion (think API push or scheduled fetch) as already structured tax fact patterns plus a skeletal memo outline.
• A Prefect flow then orchestrates a chain of LLM modules, RAG retrieval, and enrichment tasks.
• The flow returns a JSON-wrapped memo draft that a human reviewer can open immediately in our internal tools.
You’ll be free to choose the most appropriate vector store, embedding model, and validation strategy so long as everything runs on Python 3.9+ and is fully test-covered. Clean code, tight unit tests, and concise documentation are critical because future engineers will extend this work into risk identification, context-wrapper expansion, and prompt refinement.
Deliverables
- A reproducible Prefect flow (v2) in a Git repo
- Supporting modules for LLM calls, RAG retrieval, and data validation
- Unit tests that cover the core logic (pytest preferred)
- A short README explaining setup, configuration, and how to trigger the flow locally and in a Prefect deployment
Acceptance criteria
- End-to-end run produces a JSON file that mirrors the provided schema and contains a plausibly-structured memo body
- All tests pass with `pytest -q`
- Code quality meets black/flake8 standards and is type-hinted
If you thrive on architecting clean, production-ready data pipelines and know Prefect, LLM integration, and RAG inside out, I’d love to see how you’d approach this.