RefundFlow
Customer-support refund automation SaaS — users manage refund requests, AI agent (LangChain/Ollama) auto-decides based on policy, human-in-loop for edge cases.
The Refund Processing Bottleneck
Customer support refund processing is repetitive, policy-driven, and error-prone. Each refund request requires checking order history, applying refund policy, calculating amounts, and processing the refund. For e-commerce businesses processing hundreds of refunds daily, the overhead is significant.
The core challenge: build a full-stack SaaS where an AI agent handles the common cases autonomously, escalating edge cases to human operators. The system must integrate order lookup, policy enforcement, amount calculation, and refund execution — all while maintaining audit trails and supporting multi-tenant isolation.
Industry Parallels: AI-augmented workflow automation (Zapier/Retool)
Three-Tier AI-Augmented System
The system decomposes into three tiers: a Next.js frontend for refund request management, a Supabase backend for auth, data, and real-time updates, and a Python FastAPI agent service running LangChain with a local Ollama LLM for autonomous decision-making.
from langchain.agents import Tool, AgentExecutor
from supabase import create_client
tools = [
Tool(name="lookup_order",
func=lambda id: create_client(url, key)
.table("orders").select("*").eq("id", id).execute()),
Tool(name="refund_policy",
func=lambda: open("policy.txt").read()),
]
agent = AgentExecutor(agent=llm_agent, tools=tools, verbose=True)
result = agent.invoke({"input": "Refund order #ORD-4921 for $59.99"})- Next.js Frontend — App Router, TypeScript, Tailwind CSS. Customer-facing refund request portal and operator review dashboard.
- Supabase Backend — Postgres for order/refund data, Auth for multi-tenant login, Realtime for live dashboard updates and agent decision notifications.
- FastAPI Agent Service — LangChain with Ollama-powered LLM. Receives refund requests, looks up order data, applies policy via few-shot prompts, and either auto-approves/denies or escalates to a human operator.
System Architecture
Refund Request Decision Pipeline
Frontend ↔ Backend ↔ AI Decision Engine
Languages & Frameworks
Key Decisions
Auth, database, and real-time out of the box. Self-hostable with Docker for data sovereignty. Row Level Security provides multi-tenant isolation at the database level, critical for a SaaS processing refund data across different businesses.
Refund data contains PII (customer names, payment details). Local inference keeps all data on-premise. No per-request API costs at scale — important for high-volume refund processing where each request consumes LLM tokens.
LangChain works better in Python with the broader PyPI ecosystem. Async Python is natural for LLM calls with 5-30 second latency. Separates the agent concerns from the frontend — each can scale independently and be deployed on different infrastructure.
How It Was Built
Building the initial app scaffold: Next.js app router setup, Supabase schema design for orders/refunds/users, auth integration, and the agent workflow skeleton with LangChain policy prompting.
Debugging Supabase queries and agent responses. Resolved cross-origin issues, real-time subscription setup, and LangChain tool calling integration with the browser environment.
Setting up local Supabase instance via Docker Compose for development. Testing auth flows, Row Level Security policies, and end-to-end refund request processing against the local environment.
Testing the refund flow end-to-end through the browser. Validating the human-in-loop dashboard, agent decision notifications in real-time, and multi-tenant isolation for different merchant accounts.
Final integration testing of the complete stack running against the self-hosted Supabase instance. Verifying that the AI agent correctly auto-decides standard refunds and escalates edge cases to the operator dashboard.
Delivered Assets
Nine dedicated sessions produced a complete AI-augmented refund automation SaaS. The system features auth, refund request management, an AI agent using LangChain/Ollama for autonomous decision-making, and a human-in-loop dashboard for edge case handling. The architecture aligns with the broader industry trend toward AI-augmented workflow automation, similar to Zapier and Retool.
Similar Case Studies
For Enterprises
Want to Automate Your Refund Processing?
This project demonstrates how AI-augmented workflows can eliminate repetitive refund processing overhead. Let's discuss how the same patterns apply to your customer support operations.
Schedule a Strategy ConsultationFor Founders
Building AI-Augmented SaaS?
The patterns here — local LLM inference for PII safety, LangChain policy agents, human-in-loop escalation, multi-tenant Supabase — are directly applicable to your AI workflow automation products.
Discuss Your Startup's AI Vision