# Data Cleaning — Messy CRM Export → Clean, Typed Dataset

A reproducible end-to-end data-cleaning case study. You receive a "customer
orders" export from a legacy CRM in the condition real exports arrive in:
mixed date formats, duplicate rows, casing and whitespace noise, sentinel
nulls, and prices that are text. This project cleans it with pandas and
produces a **QA report generated from the actual run** — every fix is
counted, nothing is hand-typed.

The dirty data is generated by a seeded script (`make_messy.py`, seed 42),
so anyone can reproduce the exact same input and verify every number in the
QA report.

## The mess (real examples from `messy_orders.csv`)

| Problem | Example |
|---|---|
| Mixed date formats | `2024-03-01`, `03/01/2024`, `01/03/2024`, `Mar 01, 2024`, `01-Mar-2024`, `2024/03/01` |
| Ambiguous dates | `01/03/2024` — Jan 3 or Mar 1? |
| Casing noise | `joseph thomas`, `SHIPPED`, `SAN DIEGO` |
| Whitespace noise | ` san diego `, `Pending ` (trailing), double spaces |
| Sentinel nulls | `N/A`, `n/a`, `NA`, `?`, empty |
| Wrong types | `$1,234.56` as text, `(294.50)` accounting-negative, `4.0` where an int is expected, `ORD 10001` (space not dash) |
| Duplicates | 25 exact + 15 near-duplicates (same order, different casing) |

## What the pipeline does

`clean_with_pandas.py` reads everything as `dtype=str` (no silent coercion),
then:

1. Strips whitespace on all cells — counted per run.
2. Converts sentinel nulls to real nulls.
3. Normalizes `order_id` (`ORD 10001` → `ORD-10001`).
4. Title-cases names/cities, lower-cases email/payment/status.
5. Parses dates against an **explicit format list** (no `guess_date`).
   Ambiguous `X/Y/YYYY` (both parts ≤ 12) is assumed US month-first and the
   count is reported — a documented, auditable decision, not a guess.
6. Casts `quantity` to int and money columns to float (`$`/`,`/`()` handled;
   accounting parentheses → negative, i.e. credit notes).
7. **Deduplicates last** — a whitespace-padded duplicate is only an exact
   duplicate *after* normalization, so order matters. Rows with a real
   key dedupe on `order_id`; keyless rows dedupe on the full row (never on
   a null key).
8. Validates: flags rows where `|total − unit_price × quantity| > 0.02`.

It writes `clean.csv` and `QA_report.md` (row counts, every issue
found/fixed with its rule, the date-format distribution, the ambiguity
count, validation results, and a data dictionary).

## Run it

```bash
python make_messy.py messy_orders.csv          # reproduce the dirty input (seed 42)
python clean_with_pandas.py messy_orders.csv clean.csv QA_report.md
python make_previews.py                         # render before/after PNGs
```

## Result of the committed run

* **540 raw rows → 500 clean rows** (exactly the 40 injected duplicates removed).
* **163** whitespace cells stripped; **292** casing fixes; **96** sentinel
  nulls; **109** money-text values; **15** `order_id` format fixes;
  **77** ambiguous dates resolved (assumed month-first, reported);
  **21** credit-note totals normalized to negative; **21** rows flagged by
  the total check — all of them credit notes, **no true arithmetic errors**.

See `QA_report.md` for the full per-issue table.

## Before / after

`before_messy.png` and `after_clean.png` show the first 13 rows of each.
The contrast is the whole point: mixed-case cities, `29-Nov-2024` / `07/06/2024` /
`Sep 10, 2024` / `2024-12-28` on the left become a single ISO `YYYY-MM-DD` on
the right; `$1,847.64` and `(294.50)` become `1847.64` and `-294.5`.

## Why this is how I clean data in production

* **Everything is counted, not assumed.** The QA report is generated from
  the run, so a client can audit exactly what changed and how much.
* **Decisions are documented.** The month-first tie-break and the
  parentheses-are-negative rule are stated in the report, so a stakeholder
  can object to a *policy*, not discover it in a dashboard.
* **Nothing is silently dropped.** Duplicates are removed by an explicit key
  rule; quality problems are *flagged* in counts and kept for review.
* **Reproducible.** Seeded generator + `dtype=str` read means the before and
  after are both verifiable from scratch.

## Files

* `make_messy.py` — seeded generator of the 540-row dirty input
* `clean_with_pandas.py` — the cleaning + QA pipeline
* `make_previews.py` — Pillow-based before/after PNG renderer
* `messy_orders.csv` — the dirty input (540 rows)
* `clean.csv` — the cleaned output (500 rows)
* `QA_report.md` — generated QA report
* `before_messy.png` / `after_clean.png` — rendered previews