Build a multi-step agent workflow

Multi-agent document-triage workflow with human approval

Build a LangGraph workflow that triages loan-application packs through an extraction agent, a policy-check agent and a summarizer agent, pausing for human approval before any application is marked approved. You generate every input yourself — a Python script plus an LLM produces 20 synthetic application packs (applicant form, salary slip, three-month statement) rendered to PDF, so no real customer documents are needed — and the rules the policy agent applies come from a lending-policy PDF you download free from the RBI website. Persist graph state to disk so a run resumes correctly after the human responds, even across a process restart. Cap the loop at a fixed number of steps and log every transition.

20 synthetic loan-application packs (applicant form, salary slip, 3-month statement) you render to PDF from a Python + LLM generator script — no real customer documents anywhere in the project

1. Write the pack generator, render 20 synthetic PDFs, and download the public policy PDF · ~2.5h
2. Build the extraction agent node and get clean structured fields out of one pack · ~3.5h
3. Add the policy-check and summarizer nodes with conditional routing between them · ~4h
4. Add checkpointing and the human-approval interrupt, then prove a resume after a restart · ~4h
5. Enforce the step cap, capture a full run trace, and write the README and graph diagram · ~2.5h

• A LangGraph (or equivalent) state graph with at least 3 distinct agent nodes and conditional routing
• A human-in-the-loop interrupt before the final approval decision, with the workflow resuming correctly after input
• State persists across a process restart (checkpointing to disk/DB)
• A step/iteration cap enforced so the graph cannot loop indefinitely, with a logged trace of a full run

Evidence a recruiter can check

• The rendered state graph (LangGraph's own export) next to a README line per node saying what it decides and what it hands on
• A run trace showing the interrupt firing, the process killed mid-workflow, and the graph resuming from its checkpoint after approval
• A log of a run that hits the step cap and exits cleanly with a reason, instead of looping
• The generator script and the 20 synthetic packs, so a stranger can reproduce the whole run without any private data
• A short write-up of one failure mode you hit — bad extraction, model refusal, or a routing loop — and the change that fixed it

Built a three-agent LangGraph document-triage workflow with checkpointed state, a human-approval interrupt and a hard step cap — it resumes mid-run after a process restart and never reaches an approval decision without a person in the loop.