Case study · Indosat Sales & Distribution · 2026
Bots that clear the busywork in enterprise portals.
The biggest piece replaced an Excel macro: staff upload a quantity sheet, every row is matched to the customer master, rows become sales orders with their Oracle numbers, and the three files Oracle Fusion imports come out zipped. Around it, small browser bots do the clicking in the portals people used to click through by hand.

Screens: the real app running on made-up demo data. No real people, outlets or figures.
01 · Sales orders, from a sheet
Upload, check, generate.
- 1
Upload
Pick the brand, the product tab and the transaction date, then upload the quantity sheet: one row per dealer, one column per product.
- 2
Match
Each row is matched to the customer master (ship-to and sold-to) and each product to its order type.
- 3
Preview
Totals, rows skipped with the reason, warnings, and the orders about to be created; errors download as Excel.
- 4
Generate
Numbers are assigned in one locked transaction; the three import CSVs, a ZIP and a filled copy of the original macro workbook are written.
- 5
Send
The batch page offers the downloads and can email them.
02 · Try a sheet
Rows in, orders out.
A tiny made-up customer master. Edit the sheet: rows are grouped into orders and numbered the way the app does it, and anything that doesn't match is skipped with the reason.
| MPC ID Mobo | Sold To | Ship To | 9100000101SP 5GB | 9100000102SP 10GB | 9100000201Voucher 3GB | 9100009999not configured |
|---|---|---|---|---|---|---|
| MOBO70101 | ||||||
| MOBO70102 | ||||||
| MOBO70103 |
Try: change a Ship To to DLR991, or type a quantity under 9100009999.
Orders to be created · 5
2144000000001 · MOBO70101 → DLR011
SP 5GB × 300 · SP 10GB × 50
1244000000001 · MOBO70101 → DLR011
Voucher 3GB × 100
2144000000002 · MOBO70102 → DLR021
SP 5GB × 150
2144000000003 · MOBO70103 → DLR031
SP 10GB × 200
1244000000002 · MOBO70103 → DLR031
Voucher 3GB × 80
Errors (skipped) · 0
- None. Every row matched the master.
03 · Under the hood
Byte-for-byte what Oracle expects.

A forgiving reader
Finds the header in the first rows by its key column, accepts several spellings of Sold To and Ship To, reads 1.000 and 1,000 as thousands, cleans codes like 3530000023.0, and skips blanks and zeros.
Matching
Ship To → the dealer's site (Party Site ID); Sold To → the customer's billing site (Account Site ID). If a dealer sits under several customers, the row's own customer wins, with a warning.
One order per group
Rows group into one order per ID Mobo, ship-to, sold-to and order type; the same product twice is summed, with a warning.
Numbering
Identifier = circle + date + a 3-digit sequence; number = order-type code + 4 + circle code + 9 digits. Both counters reset monthly and can be set by hand.
Exact output
Fixed column counts, CRLF, UTF-8 without BOM; a test checks the files match the macro's sample byte for byte.
The macro, kept
The original .xlsm is filled at the XML level so its buttons still work; if that fails, the CSVs still go out.
04 · The bots around it
Clicking, done by software.
Small tools, each for one portal. Most are Chrome extensions that work inside the page the person is already signed in to.
Oracle Fusion
Download sales-order PDFs in bulk, schedule acknowledgement reports, cancel orders (dry-run by default, with a limit per run), add email and phone contacts to customer sites from Excel.
Mobii
Approve orders in the inbox, upload CSV changes, add transition nodes to users, edit parent values per sub-district, keep the session alive.
SalesHub
Update user data and validate IDs from Excel, through the app's accessibility tree (it's Flutter Web).
Mobile Money
Pull the transaction report one day at a time over a date range, as an extension or a Playwright tool.
Reports to WhatsApp
BigQuery → Excel master → Outlook email → WhatsApp, with a WhatsApp sender service and a small 3D office that shows each bot's status.
05 · The small stuff
The parts nobody sees.
Sessions
PBKDF2 with 240,000 rounds, a 12-hour session, five failed logins in five minutes per user and address get a 429.
Reports from billing
Billing and shipping exports become a monthly report by circle, region, branch and partner: allocation, paid and waiting payment.
Readable errors
The error workbook has summary, error and warning tabs, frozen headers and filters.
Fast enough
Customer lookups are batched 500 keys at a time; SQLite runs in WAL mode so reading never waits for a big import.
History
Every page has its own audit trail in a side panel.
Email, carefully
A test-connection button, the password never shown again, an 18 MB attachment cap and a mail log.
Bots that stop
Destructive bots run dry by default, ask before acting, cap each run and stop at the first error.
Always on
Runs on an office machine, restarted by the scheduler if it isn't listening.
06 · How it fits together
A small server and a shelf of bots.
Browser
Server-rendered pages
Upload, preview, batches, counters, customers, materials, reports, email.
Office machine
FastAPI
Login, template reader, planner, numbering, CSV and ZIP writer, report builder, mailer.
SQLite
master data, counters, batches, audit
Files
ZIPs, filled macro, error reports
Chrome + Node
The bots
Extensions inside Oracle Fusion, Mobii, SalesHub and Mobile Money; Playwright for long report pulls; a WhatsApp sender service.
07 · Tech stack
What each piece does.
FastAPI + Uvicorn
The sales-order web app.
Jinja2
Server-rendered pages.
openpyxl
Reading templates, writing reports and error files.
SQLite
Master data, counters and history.
csv, zipfile, XML
The Oracle import files and the filled macro.
Chrome extensions (MV3)
Bots inside the portals.
Playwright
Long, unattended report pulls.
Node + Baileys
The WhatsApp sender.