Ledger pilot
Link to open source: https://github.com/dwivedi-mohit/swytchcode-hackathon-build-by-mohit
LedgerPilot is an AI agent that automates billing day for a finance team: it reads inbox emails to discover invoices, classifies each one (overdue, disputed, due-soon, paid), and autonomously acts on them — chasing overdue invoices through PayPal, escalating disputes as tracked issues to Jira, logging every outcome to a Notion ops-log, and posting a final summary to Slack. Security and governance are first-class: every money-moving action is held behind a policy-enforced human-approval gate (paypal-approval in policies.json) that binds the approval to the exact request via request_hash + invoice_id (a mismatch returns 409 and cannot be bypassed); all calls are idempotent via Idempotency-Key so retries never create duplicate rows or issues; the agent runs on a scoped service identity with read/write permissions limited to the four connected providers; and the entire decision flow is streamed as a live, immutable audit trace (SSE + replay-safe dedupe) so operators can inspect, approve, or deny every step in real time. The goal is to replace a repetitive, multi-tool manual process with a transparent, policy-governed agent whose every decision is auditable and reversible.
