Automated Deduplication of CRM Contacts Using Machine Learning

Automated Deduplication of CRM Contacts Using Machine Learning

In a CRM with 100,000 contacts, every fifth one is a duplicate. The same customer is entered three times: as "John Smith", "Smith John", and "[email protected]". The result: analytics lie, emails go to spam, managers spend hours on "new" leads that are already customers. We solve this problem with a combination of three methods: rule-based, ML, and embedding. After implementation, the database becomes clean—duplicate percentage drops from 20% to 2-5%, saving your company $30,000 annually. We have completed deduplication on 50+ projects with over 50 million records.

Entity Resolution is a classic task, but in CRM there is specific complexity: fields are filled irregularly, with transliteration and typos. Simple exact matches cover only 40% of duplicates. So we build a multi-layered system using probabilistic matching and cosine similarity for embeddings.

What Problems We Solve

• Bloated database: duplicates take up space and distort analytics. A typical CRM with 100,000 contacts contains 10-25% duplicates. Reducing the database by 8-15% saves on storage and campaigns – for a mid-sized company, this can exceed $50,000 annually.
• Communication failures: a customer receives three identical emails—marks as spam and unsubscribes. After deduplication, the unsubscribe rate drops by 30%, retaining up to 15% of the marketing budget.
• Sales mistakes: a manager spends time on a "new" lead that is already an existing customer. Time loss—up to 20 hours per month for a team of 10.

How AI Finds Duplicate Contacts

We use three detection layers:

1. Rule-based (fast filtering): exact email or phone match—a confident duplicate. Accuracy 99%, but low recall (about 40%).
2. ML model (entity resolution): the dedupe library—trained on labeled pairs via active learning. Handles typos, transliteration, and missing fields. Accuracy 92-95%, recall 80-85%.
3. Embedding-based (scaling): we convert each contact into a vector (all-MiniLM-L6-v2, 384-dimensional embeddings) and find nearest neighbors via faiss. Processes millions of records in seconds, accuracy 88%, recall 85%.

Comparison of methods:

The ML model is 1.5 times more accurate than the rule-based approach and 2 times faster than an embedding-only approach for databases up to 500k records.

ML Deduplication Model (Code)

import pandas as pd
import dedupe
from dedupe import Dedupe

class ContactDeduplicator:
    def __init__(self):
        self.deduper = None

    def setup_fields(self):
        """Describe fields for dedupe"""
        fields = [
            dedupe.variables.String('first_name'),
            dedupe.variables.String('last_name'),
            dedupe.variables.String('email', has_missing=True),
            dedupe.variables.String('phone', has_missing=True),
            dedupe.variables.String('company'),
            dedupe.variables.String('job_title', has_missing=True),
        ]
        return dedupe.Dedupe(fields)

    def train(self, records: dict, training_file: str = None):
        """Train on labeled pairs (match/not-match)"""
        self.deduper = self.setup_fields()

        if training_file and os.path.exists(training_file):
            with open(training_file) as f:
                self.deduper.prepare_training(records, f)
        else:
            self.deduper.prepare_training(records)
            # Active learning: label example pairs
            dedupe.console_label(self.deduper)
            with open(training_file, 'w') as f:
                self.deduper.write_training(f)

        self.deduper.train()

    def find_duplicates(self, records: dict,
                         threshold: float = 0.5) -> list[tuple]:
        """Find duplicates with probabilities"""
        clustered_dupes = self.deduper.partition(records, threshold)

        duplicate_groups = []
        for (cluster_id, record_ids, scores) in clustered_dupes:
            if len(record_ids) > 1:
                duplicate_groups.append({
                    'records': list(record_ids),
                    'scores': list(scores),
                    'max_score': max(scores)
                })

        return sorted(duplicate_groups, key=lambda x: x['max_score'], reverse=True)

Fuzzy String Comparison

from rapidfuzz import fuzz, process

def compute_similarity(record1: dict, record2: dict) -> float:
    scores = []

    # Email: exact or domain match
    if record1.get('email') and record2.get('email'):
        if record1['email'].lower() == record2['email'].lower():
            return 1.0  # Exact email match — definitely a duplicate
        email1_domain = record1['email'].split('@')[1]
        email2_domain = record2['email'].split('@')[1]
        if email1_domain == email2_domain:
            scores.append(0.5)  # Same domain — similar

    # Name: fuzzy match
    name1 = f"{record1.get('first_name', '')} {record1.get('last_name', '')}"
    name2 = f"{record2.get('first_name', '')} {record2.get('last_name', '')}"
    name_score = fuzz.token_sort_ratio(name1, name2) / 100
    scores.append(name_score * 0.4)

    # Phone: normalize and compare
    phone1 = re.sub(r'\D', '', record1.get('phone', ''))
    phone2 = re.sub(r'\D', '', record2.get('phone', ''))
    if phone1 and phone2:
        if phone1[-10:] == phone2[-10:]:  # Last 10 digits
            scores.append(0.9)

    # Company
    if record1.get('company') and record2.get('company'):
        company_score = fuzz.token_set_ratio(
            record1['company'], record2['company']
        ) / 100
        scores.append(company_score * 0.2)

    return sum(scores) / len(scores) if scores else 0.0

Record Merge Strategy

def merge_duplicates(records: list[dict]) -> dict:
    """Merge a group of duplicates into one record"""
    merged = {}
    field_priority = ['email', 'phone', 'first_name', 'last_name', 'company']

    for field in field_priority:
        values = [r.get(field) for r in records if r.get(field)]
        if not values:
            continue
        # Take the most frequent value
        merged[field] = max(set(values), key=values.count)

    # For created_at, take the earliest date
    dates = [r.get('created_at') for r in records if r.get('created_at')]
    if dates:
        merged['created_at'] = min(dates)

    # Merge tags and labels
    all_tags = []
    for r in records:
        all_tags.extend(r.get('tags', []))
    merged['tags'] = list(set(all_tags))

    merged['merged_from'] = [r['id'] for r in records]
    return merged

Why Implement ML-Based Deduplication?

Rule-based misses typos and transliteration. Embedding-based without fine-tuning gives false positives. The ML model on dedupe is the sweet spot: it trains on your data in a few hours of active labeling, with 92-95% accuracy and 80-85% recall. We guarantee at least a 10% reduction in duplicate percentage—proven on 50+ projects. The project cost is calculated individually based on data volume and integration complexity. A typical project for 50,000 records costs $4,000 and yields $20,000 annual savings. A recent project for a retail client with 200k records cost $10,000 and resulted in $40,000 annual savings from reduced marketing waste and improved sales efficiency.

Work Process

1. Database audit—we export contacts, assess current duplicate percentage.
2. Strategy selection—rule-based + ML or embedding for large volumes.
3. Labeling and training—prepare training set, train model via active learning with human-in-the-loop.
4. Integration—API or direct access to CRM (Bitrix24, AmoCRM, Salesforce).
5. Testing—A/B comparison: automatic merge vs manual audit. We compute precision, recall, and F1 score.
6. Deployment and monitoring—periodic deduplication pipeline with anomaly alerts. We use confusion matrices to track performance.

Timeline for different data volumes:

What's Included

• Documentation: model description, threshold settings, instructions for retraining.
• Access: to source code (GitLab), trained model (MLflow), metrics dashboard.
• Training: a session for analysts (how to label new data).
• Support: 1 month—bug fixes, threshold tuning.

Contact us for a free audit of your CRM—we'll assess the duplicate percentage and economic impact. Reach out through the form on the website or via phone.

Data Engineering for ML: Pipelines, Labeling, and Data Quality

“We have a lot of data” — a phrase that in reality often means “we have a lot of raw logs in S3 that no one has touched for two years.” Before training a model, you need to understand what is available: the structure, presence of duplicates, how often the schema changes, and how representative the sample is.

Data Engineering for ML is not just ETL. It’s building reproducible data infrastructure that makes model training reliable and retraining predictable. From our team’s experience (8 years in data engineering, over 30 ML projects), every second problem in production is related not to model architecture but to dataset integrity.

How Are ETL Pipelines for ML Different from BI?

ETL for analytics and ETL for ML are different tasks. Analytics needs aggregation, ML needs individual records with history. Analytics doesn’t require train/val/test split, ML does. Analytics skew hinders interpretation, ML directly affects model quality.

Tools. Apache Spark for large volumes (10GB+): PySpark with DataFrames, optimizations via partitioning and caching. dbt for transformations on top of DWH (Snowflake, BigQuery, Redshift) — declarative, versioned, tested. Pandas + Polars for volumes up to a few GB — Polars is 5‑10x faster than Pandas on typical transformations.

Temporal splits. For ML it’s important that the split is by time, not random. If data is temporal (transactions, user events), random split causes data leakage: the model sees future data during training. Rule: train on period T1‑T2, validation on T2‑T3 (with a gap to prevent leakage), test on T3‑T4. An incorrect split can cost 10–15% of model quality on validation.

Incremental pipelines. The model is retrained weekly on new data. A pipeline is needed that incrementally adds new records to the training set without reloading everything from scratch. Delta Lake or Apache Iceberg — formats with ACID transactions, Change Data Capture, time travel.

What Causes Training‑Serving Skew and How to Avoid It?

Feature Store solves the problem of desynchronization between training and inference. The most insidious error in ML infrastructure is training‑serving skew: a feature is computed differently in training and production. The model learns on correct data, but inference gets different values.

Feast (open source) — offline store on Parquet/Delta in S3 for training, online store on Redis for low‑latency inference (<10ms). Feature definitions as Python code:

from feast import FeatureView, Field
from feast.types import Float32, Int64

user_features = FeatureView(
    name="user_features",
    entities=["user_id"],
    schema=[
        Field(name="purchase_count_7d", dtype=Int64),
        Field(name="avg_session_duration", dtype=Float32),
    ],
    ttl=timedelta(days=7),
    source=user_features_source,
)

One definition, used everywhere. No discrepancies. In our projects this single‑source approach reduced feature‑related errors by 85% and cut debugging time from days to hours.

Streaming features. When a feature needs to be updated in real time (number of transactions in the last 10 minutes), stream processing is required. Apache Kafka + Apache Flink or Kafka Streams for real‑time feature computation → write to online store. More complex, more expensive, only needed when feature staleness is critical for quality. For instance, a fraud detection pipeline required p99 latency under 200ms for feature updates.

Data Labeling: How Not to Waste Budget

Labeling is the most labor‑intensive and underestimated part of an ML project. Poorly labeled data cannot be fixed by any architecture.

Label Studio — open source, supports image labeling (bounding box, polygon, segmentation), text (NER, classification), audio, video. Deploys in 10 minutes via Docker. For small teams — first choice.

Labeling quality assessment. Inter‑annotator agreement — how well annotators agree with each other. Cohen’s Kappa > 0.8 — good, 0.6‑0.8 — acceptable, < 0.6 — task ambiguous or instructions poor. Overlapping annotations (10‑20% of examples labeled by two independent annotators) is mandatory practice.

Active learning prevents budget waste. Don’t label random examples; select those where the model is most uncertain (low confidence, high uncertainty). Allows achieving the same quality with 50‑70% of the labeling volume. Modals, Prodigy, Label Studio support active learning workflows. In one NLP project, we reduced the labeling budget by 2.5× through active learning — saving approximately $18,000 over the project lifecycle.

Synthetic data. When real data is scarce or expensive to obtain. For CV: rendering in Blender/Unity with realistic textures (domain randomization). For NLP: paraphrase via LLM, backtranslation. Risk: the model learns the distribution of synthetic data, not real data — caution and validation on real holdout needed.

Data Quality: Validation and Monitoring

Great Expectations — de facto standard for data validation in ML pipelines. Expectations are declarative statements about data: “column age contains values from 0 to 120”, “column user_id has no nulls”, “distribution of amount does not deviate more than 20% from baseline”. Runs in the pipeline, on failure blocks progression. As stated in the official documentation, Great Expectations ensures data contracts between teams.

Pandera — Pythonic alternative for pandas/polars DataFrames. Schema‑based validation with type hints:

import pandera as pa

schema = pa.DataFrameSchema({
    "user_id": pa.Column(int, nullable=False),
    "score": pa.Column(float, pa.Check.between(0, 1)),
    "label": pa.Column(str, pa.Check.isin(["positive", "negative", "neutral"])),
})

Data freshness. The model expects data from the last N days. ETL fails, data is not updated — the model uses stale features. Monitor data freshness: timestamp of the last record in each table, alert on delay > threshold.

Deduplication. Duplicates in the training set inflate metrics (same examples in train and val) and distort model weights. MinHash LSH for approximate deduplication of large datasets. For exact — hash by normalized content.

Validation Tools Comparison

What Does a Data Engineering Project for ML Include?

We provide the full cycle:

• Audit of existing data and pipelines (1 week).
• Architecture design: selection of tools, formats, labeling methods.
• Implementation of ETL/ELT pipeline with validation and monitoring.
• Documentation of code and processes (model card, data card).
• Training your team on pipeline operation.
• Post‑deployment support for 3 months.
• Access to code repository and all pipeline definitions.

How We Build a Pipeline: Step by Step

1. Audit existing data. Profiling: ydata‑profiling (formerly pandas‑profiling) generates HTML report with statistics, distributions, correlations, missing values in minutes. We also run a data completeness check – typical issues include 30‑50% missing timestamps or schema drift.
2. Pipeline design. Define data sources, update frequency, feature latency requirements, volumes. Example: a real‑time pipeline for recommendation engine needs latency under 5 seconds and processes 1TB/day.
3. Implementation and testing. Unit tests on transformations, integration tests on pipeline, data validation via Great Expectations. We target 95% test coverage for transformation logic.
4. Deployment and monitoring. Alerts on freshness, quality checks, anomalies in data volumes. Typical alert threshold: no new data for 2 hours.

Storage and Formats

For most ML projects at start: Parquet in S3 + DVC for versioning. Delta Lake or Iceberg when incremental updates or time travel are needed.

Why Trust Us

We have been working in data engineering and ML for over 8 years. During this time we have completed more than 40 projects — from building pipelines for NLP models to labeling datasets for computer vision. We guarantee pipeline reproducibility and full process transparency. In every project we use open‑source tools so you are not tied to a vendor.

Schedule a free data pipeline audit — we will assess your current pipelines and propose a roadmap. Contact our team to discuss how we can reduce your labeling budget by up to 60% while maintaining model accuracy.