LedgerLens — Revenue intelligence

ledgerlensREVENUE INTELLIGENCE

FINANCIAL DETECTIVE

Financial detective · LangGraph

Inside an investigation

See how tool calls, guardrails, verification, and your approval fit into one agent run.

01Your questionPlan, invoice, or action
02Input guardrailsBlock injection; redact PII
03Agent nodeOpenAI chooses a tool or answer

Read and draft toolsAgent ↔ ToolNode

Agentchooses a tool call

tool call result

ToolNoderuns the tool

Readload_planquery_invoicesfx_convert

Draft onlypropose_make_good_invoicepropose_credit_memopropose_plan_amendment

Results return to the model for another step. Tool calls and outcomes are logged in chat.

LoopGuardrail: caps the run at 8 agent steps, blocks repeated calls, then routes to a safe fallback.

Write path · approval requiredapply(draft) or rollback(action_id)

ApprovalPolicyGuardrail catches either write → LangGraph interrupt() pauses and shows the action in chat.

ApproveCommand(resume) continues; ToolNode runs the write.

RejectGraph ends; the write tool never runs.

Apply adds the approved draft to the sandbox. Rollback removes that action by ID. Both are written to the audit log.

GroundednessGuardrail · before the answerChecks amounts and IDs against tool results. If unsupported, the agent retries once; unsupported claims are withheld.

Your API key powers OpenAI calls and is not saved with chats.

Welcome back.