About the role
Requirements:
- 5+ years of backend Python.
- Has shipped at least one production LLM agent (tool-using, not a chat wrapper). They should be able to describe what broke and how they measured quality.
- Hands-on with Claude SDK / OpenAI Assistants / LangGraph (Claude SDK preferred your stack).
- RAG in production: chunking strategy choices, hybrid search, retrieval evaluation.
- Vector DB experience: Qdrant, pgvector, or Pinecone.
- Treats prompt engineering as an engineering discipline with eval suites, not just "vibes-based" iteration.
- Has built at least one system with feedback loops (not static RAG that's frozen on day one).
Strong nice-to-haves:
- Frappe / ERP Next experience (willingness to learn; it is fine; it's quirky but documented), Insurance, fintech, or another regulated data domain, Telegram Bot API.
- Document AI: Google Vision, Textract, or layout-aware OCR.
- Eval frameworks: Braintrust, LangSmith, or have rolled their own.