Learn AI and Data Engineering

Building an agent with LangGraph and running it: python -m extractor --batch reports/2026-Q3. 38 PDFs · 1,240 pages · ~4 h of manual work. pull revenue, margin and debt from every report. Before writing anything, I'll define the graph and. dry-run it over three documents.. from langgraph.graph import StateGraph, START, END. from langgraph.checkpoint.memory import InMemorySaver. from langgraph.types import interrupt, Command. class State(TypedDict):. document: str. figures: dict. review: bool. builder = StateGraph(State). builder.add_node("read", read_pdf). builder.add_node("extract", extract_figures). builder.add_node("review", ask_confirmation). builder.add_node("save", save_row). builder.add_edge(START, "read"). builder.add_conditional_edges("extract", route). builder.add_edge("save", END). agent = builder.compile(checkpointer=InMemorySaver()). graph compiled · 4 nodes · 1 branch · 1 interrupt. run the three test files. read → pdf.text("acme-q3.pdf"). extract → model(schema=Figures). acme-q3 · revenue 4.2M · margin 31% · debt 0.9M. read → pdf.text("borealis-q3.pdf"). borealis-q3 · no text layer · it's a scan. This one has no text. The branch sends it to OCR. instead of inventing the numbers.. route → ocr. ocr.read(dpi=300). borealis-q3 · revenue 1.8M · margin 12% · debt 2.4M. what if the model gets a figure wrong?. That's why nothing writes itself. The review node. halts the run and hands it back to you.. def ask_confirmation(state):. return interrupt({"figures": state["figures"]}). ⏸ interrupt · waiting for human confirmation. helios-q3 · margin 91% · confidence 0.41. That 91% doesn't match the sector. Flagged, not saved.. fix it to 19 and carry on. agent.invoke(Command(resume={"margin": 19})). resumed from checkpoint · step 3 of 4. good, run it over all 38. agent.stream(batch, config). 38/38 · 35 automatic · 3 flagged for review. 6 min 12 s · 0.42 USD in tokens. Four hours of work in six minutes, and the three. doubtful cases are flagged, not buried.