Skip to content
Selected work

Case study 02 / 02

02 · Automation System

GONAMI

A WhatsApp-based finance automation system with conversational workflows and AI-assisted processing.

GONAMI brings finance workflows into WhatsApp. Requests, confirmations, and document handling happen inside the conversation, with a Go service doing the heavy lifting behind the scenes.

01

Overview

An automation system that turns WhatsApp conversations into structured finance workflows. A Go service ingests events, maintains the state of each conversation, and drives the steps of the workflow until it is complete.

  1. 01Conversational finance workflows handled in WhatsApp
  2. 02AI-assisted extraction and structuring of message content
  3. 03Attachment and document handling over S3-compatible storage
02

Problem

Finance processes usually live across chat, documents, and spreadsheets. Manual back-and-forth is slow, nothing is auditable in one place, and there is no reliable record of what actually happened in a conversation.

03

Solution

A Go service integrates with the WhatsApp API to ingest and reply to conversations. PostgreSQL stores conversation and workflow state, MinIO provides S3-compatible storage for attachments, and Gemini assists with extracting and structuring content from messages.

04

Architecture

  1. 01Go service handling WhatsApp event ingestion and outbound messages
  2. 02Conversation-scoped state persisted in PostgreSQL
  3. 03MinIO / S3-compatible storage for attachments and documents
  4. 04Gemini integration for content extraction and AI-assisted processing
  5. 05Worker-based execution for slow or retryable workflow steps
05

Implementation

Incoming WhatsApp events are normalized and routed based on conversation context. Steps that need time, external calls, or human confirmation are executed asynchronously, and AI-assisted extraction feeds a structured pipeline with deterministic fallbacks.

  1. 01Event ingestion from the WhatsApp API with deduplication and ordering
  2. 02Workflow state machine persisted per conversation
  3. 03AI-assisted parsing of messages with rule-based fallbacks
06

Challenges

Keeping long-lived conversational flows reliable across retries, handling duplicate and out-of-order events from the messaging layer, and combining AI output with predictable business rules rather than treating it as a black box.

07

Outcome

An end-to-end pipeline that moves finance conversations from manual chat handling into structured, recoverable processing.