← All projects

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.

Where
Independent project · 2025–present · in production
My role
Sole developer: design, code, deployment and operations
Stack
Gemini 3.1 Flash-Lite (tool calling; Gemini 3.8 Flash since November 2026), FastAPI, SQLite, Google Sheets, Google Cloud Vision OCR, Kakao Local API, RTPay bank-deposit alerts, Next.js dashboard, Sentry · Oracle Cloud

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

The system

Thai Fruit Garden architecture: Messenger customers and shop staff; a FastAPI app on Oracle Cloud with the Messenger webhook, the Gemini agent (bot mode replies automatically, observe mode never replies), five agent tools including a Kakao address check, a three-layer order check, Cloud Vision OCR and a receipt check matched against RTPay bank deposits; SQLite and Google Sheets for data; a Next.js admin dashboard for staff.
Each shop has its own Facebook Page and catalog and its own mode: bot (the agent replies automatically and pauses while staff take over a chat), observe (it never replies and turns staff chats into orders) or off. Every order goes through three checks before it is saved. Transfer receipts are matched against the bank deposits that RTPay reports from the shop's account. Click to enlarge.

What I did

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.

5,964orders captured silently from staff chats1 Jun – 10 Jul 2026, in observe mode
66%of shipped orders captured491 of 744 customer-days, against the courier's waybills (30 Jun – 2 Jul), up from 29–30% before the fixes of 29 June; most of the gap is orders placed outside Messenger
52%of eligible transfer receipts verified with no human7,411 of 14,260 matched to RTPay deposits; 5 suspected forgeries caught. Of 16,177 receipts in all, 908 went to manual bank queues by design, 733 were re-sent copies and 276 were not receipts
49,859bank deposits received from RTPay18 May – 18 Nov 2026
88%of addresses verified with Kakao4,220 of 4,776 address-like inputs; 730 more inputs were not addresses
US$0.004Gemini cost per message after tuningAbout US$13.64 a day; 96,197 messages and images processed for US$305 at list prices, US$129 of it in a 4-day model trial
2.7 smedian LLM response timeEnd to end per message, tool calls included; 2.0 s for images; 90th percentile 9.1 s (text) and 5.4 s (images)
TFG admin dashboard, Analytics, Receipts tab for Sea Sea Market, 17 May to 11 July 2026: auto-verified 7,411, staff review 6,842, suspected forgery 5, bad format 276.
The admin dashboard's receipt summary for Sea Sea Market, 17 May – 11 Jul 2026. Click to enlarge.

Example

A simplified order in Korean. Names, address and phone number are made up.

Illustration only: the flow follows the real agent, but the wording is simplified and no real customer data is shown.
← Voice phishing detection Next: Copyright metadata extraction →