UnifyData Python

UnifyData does the month-end closing of employee purchases for a group of companies. Employees buy on credit at the group's own stores, and Linx AutoSystem, the stores' management software, records each purchase. At the end of the month someone has to turn those Linx reports into a payroll deduction sheet for each company. When Linx and the HR spreadsheets share no common code, the only thing linking them is the name, and names are written in many ways: BAPTISTA in one place, BATISTA in the other; DEBORA without the accent; a surname cut short.

UnifyData matches people by name and asks a person only about what cannot be decided alone. Names that differ only by accent, case or spaces are linked automatically; similar names, repeated names and people who are not in HR become review cards; the result is the deduction sheet, one sheet per company.

Built with Python and Streamlit, it runs on your own computer. The interface is in Portuguese.

Try it in two commands (below), then click Experimentar com arquivos de exemplo (Try with sample files): made-up files, nothing is saved. Or try the web version right now, in the browser, with nothing to install.

This README is also available in Portuguese.

In 30 seconds

1. Install and start it (a minute or two the first time):

   pip install -r requirements.txt
   streamlit run unifydata.py

   On Windows, Iniciar_UnifyData.bat does both with a double click.

2. Click Experimentar com arquivos de exemplo. Three HR spreadsheets and two Linx reports load. DEBORA (Linx) is linked to DÉBORA (HR) on its own, and five cards show up in the Divergências (discrepancies) step: one similar name (Rafael Cunha Baptista in Linx, Batista in HR, 98% alike) and four people who are in no HR spreadsheet.

3. Approve the link, type a CPF (the Brazilian taxpayer ID) and a company, or ignore. When no card is left, the closing is ready, with Baixar Excel, CSV and Imprimir ou PDF (Print or PDF).

Screenshots

Taken from the sample files.

Name review, dark theme	Closing ready, light theme
Start screen, dark theme	Closing ready, dark theme

On a phone, light theme	On a phone, dark theme

What comes out

• Excel: one sheet per company with every employee in HR; those with no consumption this month show up in gray, so the sheet doubles as the payroll list. Frozen header, currency format, formatted CPF, a totals row and sheet names that are always valid in Excel. Each sheet is set to print on A4, fitting the page width.
• CSV: ; separators and decimal commas, the way Excel in Portuguese opens it, with a company column.
• Report to print or save as PDF: A4, landscape when there are three stores or more, each company kept whole on a page when it fits, and numbered pages. Ignored people are listed at the bottom, with the amount.

The files are named after the month of the due dates (Fechamento_UnifyData_2026-09.xlsx). In the sample mode, the name ends in _DEMO, and the Excel and the report say the data is made up.

What it does

Reading

• HR spreadsheets: one per company, .xls, .xlsx or .csv, with name and CPF columns, wherever the header row is. The file name becomes the company name; a workbook with several sheets becomes Company - Sheet, unless the sheet has Excel's default name (Plan1, Planilha1...).
• Linx reports: the Pendências por responsável (pending by person) CSV, one per store. Two values per person and store, overdue and total, and the due-date period.

Matching, from safest to most careful

1. Same name once accents, case and spaces are removed, and only one person in HR with it: linked automatically (the screen says how many).
2. Similar name (85% or more with token_sort_ratio, which also catches swapped word order): a card suggests the link. Approve it, or say it is someone else by typing their CPF; the card shows whose CPF it is before you confirm.
3. Name repeated in HR (namesakes, or one person in two spreadsheets): the card asks which row to charge, so nobody is charged twice. It is never chosen automatically.
4. Not in HR: type the CPF. If it belongs to someone in HR, the consumption goes to that person; if not, a record is created in the chosen company (an existing one or a new one). Or ignore it.

Each card can be ignored; ignored people and their amounts are listed on the result and in the report.

Phone and themes

• Light and dark themes, following the system, switchable from the ⋮ menu. The custom parts follow the theme even when it changes without a reload.
• A sidebar with the four steps and their status, and the AFN Systems signature; on a phone, cards and buttons take the full width.

Privacy and security

• Runs on your computer, for your computer only. The server listens on localhost, so other machines on the network cannot open it. The files stay in the memory of your Streamlit session and are not written anywhere; closing the session discards them. Streamlit's usage statistics are turned off, and the fonts are served by the app itself (static/fonts): nothing is fetched from Google Fonts.
• The sample mode loads the made-up files from exemplos/; it does not touch anything else.
• Text from the files is never trusted: names and file names are escaped before they become HTML, and CSV cells that start with =, +, - or @ get a leading apostrophe, so a spreadsheet cannot run a formula hidden in a name.

How it is built

Part	Technology
Interface	Streamlit 1.57, with its theme config and a little CSS
Data	pandas for reading and consolidating; xlrd for legacy .xls
Names	thefuzz (token_sort_ratio) for similarity
Writing	openpyxl for the .xlsx; the CSV and the report are made by the app
Fonts	Instrument Sans for the interface; JetBrains Mono for labels and the signature, hosted in the project (SIL OFL 1.1)
Tests	unittest, on every push (GitHub Actions)

Some decisions behind it:

• The rules live apart from the screen. nucleo.py has pure functions for reading, matching, consolidating and writing the Excel and the CSV; unifydata.py only draws the screen and calls them. The tests run the rules against the same sample files, without Streamlit.
• Automatic only when the match is unambiguous. A link is made alone only when the name is the same once accents, case and spaces are removed, and HR has a single spelling of it; everything else waits for a click.
• Namesakes are protected. When a name appears more than once in HR, the consumption goes to the chosen row only, and the other rows stay at zero. Consumption that a CPF has already tied to one person stays with that person, whatever is chosen later for the name.
• Clicks are safe. Every action is a callback that runs before the next redraw, so a double click on a button that already did its job does nothing.

Known limitations

• It depends on the layout of the Linx Pendências por responsável CSV.
• Matching by name needs a person to confirm the doubtful cases; that is by design.
• Nothing is kept between closings: the HR spreadsheets are read again each month. (The web version keeps a register and matches by employee code instead.)
• It needs Python on the computer; there is no hosted version.
• Some parts belong to Streamlit itself: the ⋮ menu is in English, and its Print prints the whole page (the Imprimir ou PDF button prints only the report).
• Tested in Chrome and Edge, on a computer and at phone width. The interface is in Portuguese only.

Running it

Python 3.10 or newer. On every push, CI runs the tests and compiles the app on Python 3.10, 3.12 and 3.14:

pip install -r requirements.txt
streamlit run unifydata.py

The app opens at http://localhost:8501; http://localhost:8501/?demo=1 opens straight into the sample files. On Windows, Iniciar_UnifyData.bat installs what is missing on the first run, writes an empty Streamlit credentials file (so Streamlit does not ask for an email) and opens the app.

To run the tests:

python -m unittest discover -s tests

Project structure

unifydata.py            the Streamlit app: steps, review cards, result and exports
nucleo.py               pure rules: reading, name matching, consolidation, Excel, CSV
requirements.txt        dependencies
CHANGELOG.md            what changed in each version
Iniciar_UnifyData.bat   Windows launcher
.streamlit/config.toml  light and dark themes, fonts, toolbar, telemetry off
assets/favicon.png      the UnifyData mark
static/fonts/           Instrument Sans and JetBrains Mono (woff2), with their licenses
tests/                  tests for the rules, with the sample files
exemplos/               made-up files in the Linx and HR formats
.github/workflows/      runs the tests on every push
docs/telas/             images for this README (revisao.gif included) and the social preview (og.png)

Compared with the web version

	UnifyData Web	UnifyData Python
Runs	In the browser, nothing to install	On the computer, with Python
Inputs	Linx accounts receivable PDF, HR spreadsheets, Linx CSVs	HR spreadsheets and Linx CSVs
Matching	By employee code (exact)	By name (fuzzy, with review)
Namesakes	Not an issue, the code tells them apart	A card asks which row to charge
Employee data	A register saved in the browser	HR spreadsheets read every month
Excel	People with consumption; totals as SUM formulas	Every HR employee (no consumption in gray)
Best for	Linx and HR share the employee code	Systems with no common code

License

MIT. Made by Alyssom Fernandes, AFN Systems.