Demo Case Study: CRM Cleanup — 5,000 Messy Contacts to a Deduped, Standardized List

Demonstration project. Every name, company, email and phone number in this folder is synthetic — generated to reproduce the defects we find in real CRM exports. No real contact data was used or touched. We never present demo work as client work; this exists so you can inspect exactly what you'd get.

Before: contacts_raw.csv (5,000 rows)

A contact export the way they actually arrive:

• ~560 duplicate records — some exact copies (export glitches), some the sneaky kind: same person entered twice with a different phone format, a missing title, Fairwind Roofing Inc. vs FAIRWIND ROOFING INC. , one with an email and one without
• Names and companies in random casing with stray whitespace
• 6+ phone formats ((555) 200-1234, 555.200.1234, 5552001234, +1 555 200 1234) plus invalid entries (call office, 555-12)
• States as NY / New York / ny / blank
• Lifecycle stages in 14 spellings of what should be 4 values
• Broken emails (name(at)domain.com), uppercase emails, padded emails
• Missing titles (24%), phones (17%), states (11%)

After: contacts_clean.csv

Exact numbers are in cleanup_metrics.md (auto-generated by the pipeline — we don't hand-write results). Headlines from this run:

• 734 duplicates merged (277 exact, 285 same-email, 172 fuzzy name+company) with a survivorship rule: the most complete record wins, and its duplicates donate their non-empty fields — merging adds data, never destroys it
• Every phone in one format, invalid numbers cleared and counted
• States → USPS codes; lifecycle stages → a 4-value picklist (Lead / MQL / Customer / Churned) so segmentation and reporting actually work
• Emails validated with an Email Valid audit column (YES / NO / MISSING) — nothing silently deleted
• Field completeness up despite removing 700+ rows

Process (what we'd do for you)

1. Scope (human): which CRM, how many records, what "clean" means for you (merge rules, picklists, required fields). You get a written spec before we touch anything.
2. Dry run (AI agents): the full pipeline runs on an export first — you see the metrics report and a sample of proposed merges before anything changes in your CRM.
3. Apply (AI agents, inside your account): executed in your own CRM (HubSpot, Pipedrive, Zoho, Salesforce) via native merge tools/APIs, so history and activity records are preserved. Any enrichment happens within your accounts and your data — we don't sell, pool, or retain client data.
4. QA (human): a data engineer samples merges, checks edge cases (two people sharing an office phone are not a duplicate), and signs the metrics report.
5. Deliver: cleaned CRM, before/after metrics report, and the rule set — so the next cleanup is cheaper, or so a monthly Data Care plan keeps it clean.

Files in this folder

• contacts_raw.csv — messy input (5,000 rows, synthetic)
• contacts_clean.csv — cleaned output (+ audit column)
• cleanup_metrics.md — the auto-generated before/after report
• clean_crm.py — the pipeline (readable, auditable)
• generate_crm_data.py — how the synthetic mess was made

Built by Modular Enrichment. AI-assisted delivery, human-QA'd — openly.