Mid-level Python Developer / AI Engineer Contractor
Budget: $20 – $25 USD
I’m moving our income-tax practice toward full automation and need an experienced developer to take the first production block from concept to running code. The piece you’ll tackle is the drafting engine: given structured tax fact patterns and memo outlines in strict JSON, the system should orchestrate LLM + RAG components, enrich the content, and return a clean, review-ready memo draft—also wrapped in JSON.
The orchestration layer must run in Prefect and allow easy chaining with future blocks (risk-checking, context expansion, wrapper improvements). I already have preliminary prompts and retrieval sources; your job is to wire them together, build the enrichment tasks, and expose a simple API endpoint so my team can drop a JSON file in and receive the drafted memo back.
Key expectations
• Use Prefect for flow control and retry logic
• Integrate an LLM of your choice (OpenAI, Anthropic, etc.) plus RAG retrieval against my knowledge base (Postgres + pgvector)
• Preserve every incoming JSON field, adding only the generated memo content and metadata
• Keep the codebase modular so later blocks can plug in without refactor-pain
I’ll test by feeding sample returns and verifying that:
1. The flow executes end-to-end without manual steps.
2. Output JSON validates against the provided schema.
3. The draft reads coherently and follows my outline sections exactly.
If you’ve built similar Prefect or tax-domain NLP pipelines, that’s a big plus—please mention it when you bid along with an estimated timeline for this first block.
The orchestration layer must run in Prefect and allow easy chaining with future blocks (risk-checking, context expansion, wrapper improvements). I already have preliminary prompts and retrieval sources; your job is to wire them together, build the enrichment tasks, and expose a simple API endpoint so my team can drop a JSON file in and receive the drafted memo back.
Key expectations
• Use Prefect for flow control and retry logic
• Integrate an LLM of your choice (OpenAI, Anthropic, etc.) plus RAG retrieval against my knowledge base (Postgres + pgvector)
• Preserve every incoming JSON field, adding only the generated memo content and metadata
• Keep the codebase modular so later blocks can plug in without refactor-pain
I’ll test by feeding sample returns and verifying that:
1. The flow executes end-to-end without manual steps.
2. Output JSON validates against the provided schema.
3. The draft reads coherently and follows my outline sections exactly.
If you’ve built similar Prefect or tax-domain NLP pipelines, that’s a big plus—please mention it when you bid along with an estimated timeline for this first block.