Agentic AI
WarungFlow — Autonomous Finance Ops for Micro-Businesses
An LLM-agent system that reads a warung's WhatsApp-style orders, reconciles them against QRIS and bank transfers with fuzzy name-matching, and reports cashflow and a business-health score — winner of Best Payment Use Case at OpenClaw Agenthon Indonesia 2026.
- 🏆
- Best Payment Use Case · OpenClaw 2026
- 20
- Tools in the agent registry
- ≤24
- Steps per orchestration loop
- 0–100
- Business-health score
The problem
Millions of Indonesian warung (micro-businesses) run entirely on WhatsApp orders and a mix of QRIS and bank transfers — and their bookkeeping is a mental note or a paper ledger. The hard, unglamorous part isn’t taking the order; it’s reconciling it: matching “Bu Sri ordered 3 × Rp15.000” against a bank line that reads TRANSFER SRI WAHYUNI 45000 and knowing the day balanced.
WarungFlow is an autonomous agent that does exactly that loop — read, reconcile, report — so the owner doesn’t have to.
Agentic architecture
Rather than one giant prompt, WarungFlow is built as a tool-using agent: an orchestrator runs a bounded reasoning loop (capped at 24 steps to keep it terminating and debuggable) over a registry of ~20 tools — parse an order message, look up a customer, query payments, post a reconciliation, compute a metric, draft a reminder. The LLM decides which tool to call next; the tools do the deterministic work. That split — probabilistic planning over deterministic actions — is what keeps an agent both flexible and trustworthy with money.
Payment reconciliation
The core capability is matching orders to incoming money across QRIS and bank transfers, integrated through the DOKU payment gateway (sandbox). Because Indonesian transfer descriptions are noisy and names rarely match exactly, reconciliation leans on fuzzy name-matching to link a payment to the right order and customer — the piece the judges recognized with Best Payment Use Case.
Financial intelligence
On top of reconciliation, the agent computes cashflow and a 0–100 business-health score, and generates plain-language reminders (chase an unpaid order, flag a slow week) — turning raw message-and-transfer noise into something an owner can act on in seconds.
Reflection
WarungFlow is where I learned to build with agents responsibly: bound the loop, put the money-touching logic in deterministic tools, and let the model plan rather than execute. The award was nice; the more durable lesson was how to make an LLM system you’d actually trust near someone’s cash.