Thai Fruit Garden: a multilingual ordering agent
An LLM agent for shops that sell on Facebook Messenger in Korea. In bot mode it answers customers in Korean, Thai, Vietnamese and English; in observe mode it never replies and silently turns the staff's chats into orders and checks bank transfers. It runs in production for Sea Sea Market, a Thai grocery shop in Korea, in observe mode. I designed, built and run it on my own.
Demo and evidence
A real conversation in bot mode
Screenshots from a conversation on the Thai Fruit Garden test page on 11 May 2026, the one page that ran in bot mode. Every blue message on the right was written by the agent; the customer wrote in Thai. Personal details are blurred. Click a screenshot to enlarge it.






The problem
- Orders arrive as chats: customers order on Messenger in four languages, in free form, often with photos instead of text.
- Repeated questions: staff answer the same questions about products and prices again and again.
- Copying by hand: every order is copied into a spreadsheet, and every bank transfer is checked against a receipt screenshot.
- Hard parts for an agent: products and prices it cannot guess, Korean delivery addresses, and fake or reused transfer receipts.
- Staff stay in charge: the agent has to work next to the staff, not replace them.
The system

What I did
- Tools instead of long prompts. The agent looks up products, prices and past orders through tool calls rather than relying on a prompt that lists them, so answers about products and prices come from the data, not from the model's memory.
- Korean address checks. Delivery addresses are checked with the Kakao Local API, which returns the official road-name address and postcode for the order. If an address can't be verified, the order still goes through and staff check it by hand.
- Observe mode. The agent never replies; it reads the conversation between staff and customer and extracts the name, address, products and phone number into the order database and Google Sheets, so staff no longer copy orders into the spreadsheet by hand. Orders it is unsure about are flagged for staff. This is the mode the shops used in production.
- Payment checks against the bank. RTPay sends an alert for every deposit into the shop's account. When a customer sends a transfer screenshot, Gemini reads the amount, the time and the bank from the image itself, helped by the Cloud Vision OCR text, and the receipt is matched to a deposit by amount (allowing for a transfer fee), receiving bank and time (±5 minutes). A single match is verified automatically; anything ambiguous goes to staff, and duplicates and suspected forgeries are flagged. Receipts for two accounts that have no deposit feed go straight to staff queues.
- Three checks on every order. A rule-based validator, a grounding check (is each product really in the conversation?) and a Gemini critic that reads the order summary against the chat. In observe mode they flag orders for staff review rather than blocking them.
- Photos of addresses. Other pictures, such as address labels, are read with Cloud Vision OCR for the name, phone number and address.
- Several shops, one system. Each shop has its own Facebook Page, catalog and mode, and the staff work from a Next.js dashboard: review queue, orders, receipts, deposits, catalog import and reports.
- Reliability. Failed LLM calls go to a retry queue with an outage watchdog (4,953 failed calls were recovered after Gemini outages), errors are reported to Sentry, and the database is backed up every night.
Results
Sea Sea Market used the agent in observe mode: staff answered customers, and the agent built the orders and checked the payments in the background. Numbers from the production database for the capture period, 1 June to 10 July 2026; the capture rate is checked against the shop's own courier waybills.

Example
A simplified order in Korean. Names, address and phone number are made up.
- Customer망고스틴 2kg 주문하고 싶어요. 서울 동작구로 보내주세요.
- → catalog search: "망고스틴" · pricing: 2 kg
- Agent망고스틴 2kg 상자 주문 가능합니다. 받으실 분 성함, 전화번호, 상세 주소를 알려주세요.
- Customer홍길동, 010-0000-0000, 동작구 가상로 12, 101동 202호
- → address check (Kakao Local API): road address and postcode found
- Agent주문 내용 확인드립니다: 망고스틴 2kg, 홍길동 님, 서울 동작구 가상로 12, 101동 202호. 맞으면 "네"라고 답해주세요.
- → order saved: database + Google Sheets