Phase One - Designing a New Data & Email Architecture

Objective: turn messy, flat lists into a clean, structured, relational data model the CRM can actually use, and build a sending setup that protects TGV Lyria’s domain reputation.

From flat files to a controlled staging area

Rather than importing everything into HubSpot and hoping for the best, Cashmyrr introduced a staging database using SQL and MongoDB. Data from CSV/Excel files (FR, DE, EN) was consolidated there first. MongoDB aggregation pipelines (such as $project, $unionWith, $lookup) were used to:

• normalise fields: turn all opt-ins into a single boolean "true/false" standardise dates

• dynamically exclude emails found in dedicated "complaints" or "npai" collections

• Remove all phone numbers in email (0606060606@icloud.com)/invalid format emails that could cause deliverability issues

Only once data had passed through this staging and cleaning layer did it move on to HubSpot.

Smart upserts instead of blind imports

To avoid overwriting good data already present in HubSpot, Cashmyrr built an n8n automation workflow acting as an ingestion controller:

• The workflow queries HubSpot in batches via API.

• For each email, it first checks whether the contact already exists.

• It then updates only missing fields or creates a new contact if needed.

The result: TGV Lyria ends up with a consistent, enriched CRM — not an overwritten one.

Protecting domain reputation with a dedicated subdomain

Bulk campaigns were also re-anchored on a safer email infrastructure:

• Instead of sending from the main domain, Cashmyrr configured a dedicated subdomain: newsletter.lyria.com

• Strong authentication (DKIM, SPF, DMARC) was set up to maximise inbox placement.

• Native bi-directional sync ensured that hard bounces and unsubscribes from Brevo flowed back into HubSpot approximately every six hours, keeping contact statuses aligned.

• Also configured workflows in Brevo that remove hard/soft-bounce to avoid issues in first deliveries

Phase Two - Cleaning the Legacy Database

Once the new architecture was in place, the next issue was data quality inside HubSpot itself.

Years of imports had created a “fuzzy duplicate” problem that simple email-based deduplication couldn’t solve:

• jean.dupont@company.com vs j.dupont@company.fr

• Typos or alternate spellings of first/last names

• Mixed B2B / B2C profiles

With B2B and B2C data intermingled, the deduplication work required cleaning both populations and applying arbitration rules—then, separately, restructuring B2B company accounts.

Clustering contacts before deciding what to merge

Cashmyrr started with deterministic clustering — grouping potential duplicates based on different keys:

1. Exact email address

2. “Domain + last name”

3. “Phone number + last name”

This created focused groups of contacts likely to refer to the same person or company, without merging anything yet.

Adding AI to handle the “grey zone”

For ambiguous pairs — those with an initial similarity between 0.60 and 0.84 — Cashmyrr used OpenAIGPT to analyse semantic similarity:

• Does this look like a nickname vs full name?

• Are there typical typos between the two strings?

• Are first and last names swapped or abbreviated?

GPT produced a confidence score, and only pairs reaching ≥ 0.85 were merged automatically. Everything below remained available for manual review.

“Smart” merge rules instead of destructive merges

When a merge did happen, it followed precise, field-by-field rules:

• keep the most valid email

• keep the most advanced lifecycle stage (Customer > Opportunity > Lead)

• preserve the existing owner when it made sense

This allowed TGV Lyria to reduce noise in the database without losing valuable information.

Restructuring B2B accounts with parent / child companies

For B2B data, Cashmyrr also introduced a Parent / Child model:

• Companies sharing the same root_domain (for example hermes.com) were identified.

• A script then designated a “parent” company (often the head office) and linked local entities as child companies.

This gives teams a global view of key accounts while keeping local contacts attached to their relevant subsidiaries.

Phase Three - Fixing the Data at the Source

Cleaning legacy data is one thing. Stopping pollution at the source is what makes it sustainable.

The problem: uncontrolled web inputs

Several “home-made” sources were injecting low-quality data directly into HubSpot:

• Pure HTML forms

• Search bars and on-site features that could create contacts automatically

This led to thousands of “Organic Leads” with almost no usable information or consent. At the same time, marketing cookies were being set before users had accepted them, raising GDPR concerns.

Locking down HubSpot’s open doors

Cashmyrr tackled this in two moves:

1. API lockdown

   • The HubSpot feature that allows contact creation from “Non-HubSpot Forms” was disabled.

   • Only controlled, intentional touchpoints could now create new contacts.

2. Conditional tracking via CookieFirst

   • The Google Tag Manager integration was rewritten.

   • The HubSpot tracking script and the hubspotutk cookie (which ties web visits to CRM contacts) now only fire after explicit consent via the CookieFirst CMP.

Automatic segmentation from day one

New contacts are automatically segmented by source and intent through HubSpot workflows:

• If the source is “Newsletter form” → the contact is tagged as a B2C target

• If the source is Salesforce or a specific Owner → the contact is tagged as a B2B target

This means cleaner reporting and more relevant campaigns from the moment a new contact enters the database.