Building a Back Office SaaS with AI: Invoice, Email Reminders & Scheduling in 3 Days

Tags: #SaaS #FastAPI #AI #BackOfficeAutomation #PostgreSQL #Python #BuildLog

---

Introduction

"Can we automate small business back office operations with AI?" — that's the question that started this project.
We call it AI Back Office Pack. Here's how we built Phases 1–3.

---

What We Built

AI Back Office Pack is a modular back office automation SaaS for small businesses. Pick the modules you need and combine them.

Module	Function
Module 1	Invoice Management
Module 2	Email Reminders
Module 3	Scheduling & Meeting Room Management
Module 4	HR & Attendance (in progress)
Module 5	Document Management (planned)

---

Infrastructure

infra server (internal)
├── API Server        port 4000  (FastAPI + Python)
├── Dashboard         port 4001  (React 18 + Tailwind)
├── PostgreSQL        port 5433  (persistence)
└── MinIO             port 9000/9001  (file storage)

Key decision: everything behind Nginx reverse proxy with relative paths.
The dashboard calls /api/ instead of hardcoded localhost:4000, so it works from any machine without changing config.

---

Module 1: Invoice Management

The first module we built. Core features: CRUD + status workflow (draft → approved → sent).

Design choices:

• PDFs stored in MinIO (S3-compatible), metadata in PostgreSQL
• tenant_id on every table for multi-tenant isolation
• Multi-currency schema from day one

We hit a currency display bug after launch — everything showed $ instead of ¥. Currency fix became Phase 4's first task.

---

Module 2: Email Reminders

Automatically send payment reminder emails for invoices approaching their due date.

Invoice due date → 3 days before / 1 day before / day-of → auto email

Implementation notes:

• Scheduler: Python APScheduler, jobs registered on startup
• Template engine generates body (company name, amount, due date)
• Sent history in DB to prevent duplicate sends

Email reminders had the highest demo impact — users feel automation when an email arrives in their inbox.

---

Module 3: Scheduling & Meeting Room Management

Meeting room booking with external calendar sync (Google Calendar / Outlook).

Hardest part: time overlap checking.

SELECT * FROM room_bookings
WHERE room_id = $1
  AND start_time < $2
  AND end_time > $3;

Simple query, but timezone handling took several iterations. The server was running in JST while storing timestamps in UTC — classic timezone mix. Fixed by normalizing everything to UTC at the DB layer.

---

Tests: 25/25 PASS

All modules, all tests passing.

Module 1 (Invoice)       ✅  9/9
Module 2 (Reminder)      ✅  7/7
Module 3 (Schedule)      ✅  9/9
─────────────────────────────────
Total                    ✅  25/25

We followed a TDD-adjacent flow: API schema → tests → implementation. When the schema is fixed upfront, AI-generated implementations stay on track.

---

The Dashboard Headache

The React dashboard had localhost:4000 hardcoded for all API calls. When Linou opened it in a browser on a different PC — no connection.

Fix: change all API calls to relative paths → route /api/ through Nginx to FastAPI.

location /api/ {
    proxy_pass http://127.0.0.1:4000/;
}

After that fix: "Login works! Invoices are showing!" — Phase 3 done.

---

Lessons Learned

1. Schema first, everything else follows
Fixing the DB schema and API schema before writing code makes AI implementation prompts simple: "implement CRUD for this schema." Clear spec = predictable output.

2. Tests are the spec
Writing tests first makes "what are we building" concrete. It doesn't matter how the implementation is written as long as tests pass — ideal for AI/human collaboration.

3. The localhost trap
Hardcoded localhost will happen. Add "test from a different machine" to your checklist from day one.

---

What's Next

• Module 4: HR & Attendance (awaiting requirements confirmation)
• Currency bug fix: $ → ¥
• Stripe payment integration (Phase 4)

---

Published: 2026-03-05 / Author: Jack