The Karg platform

Clean, reconcile and review your organization's data in one place. Karg profiles every column, writes a cleaning plan, and fixes the data in three layers, with a person deciding anything it isn't sure of.

Request platform accessOr have our team do it

What Karg fixes

The problems that make spreadsheets, CRM exports and databases hard to trust, and that teams without a data engineer end up fixing by hand.

Duplicate records

Exact duplicates are removed automatically; likely duplicates are grouped and sent to you with Karg's reasons.

Inconsistent values

State names, company suffixes, categories and yes/no answers written one way across the file.

Formatting problems

Dates, phone numbers, emails, currencies and units in one consistent format.

Records across systems

Entity resolution links the same customer, supplier or product when each system spells it differently.

Data-quality exceptions

Values that look wrong for their column are flagged, and anything Karg can't resolve is listed, not guessed.

Inconsistent State Names

Geo normalization

CA | Calif | CA - USA

->

California

Rule + AI layers unify mixed location tokens.

Duplicate Customers

Entity resolution

John Smith | J. Smith

->

John Smith

Identity variants collapse into one canonical record.

Broken Phone Numbers

Format standardization

5551234567 | (555)123-4567

->

+1-555-123-4567

Deterministic formatting yields analysis-ready phone fields.

Mixed Date Formats

Date harmonization

03/02/24 | March 2 2024

->

2024-03-02

Multiple date styles normalize to ISO output.

How the platform works

Six steps from a messy source to clean, analysis-ready data.

  1. Bring your data

    Upload CSV, Excel or JSON files, or connect a database like Postgres, MySQL, Snowflake or BigQuery, an S3 bucket, or a CRM like Salesforce or HubSpot.

  2. Profile

    Karg reads every column: what it holds, which formats it uses, and what looks wrong.

  3. Plan

    It writes a cleaning plan and asks you a few questions. You approve the plan before anything changes.

  4. Clean

    Rules, specialized machine learning models and AI database agents work through the data, each value going only as far as it needs.

  5. Review

    Fixes below your confidence threshold wait in a review queue. Approve or reject each with one key.

  6. Export

    Download the clean data as a file, or send it straight back to the database it came from.

Why three layers

Most rows never need an AI model. Rules and specialized models fix them for almost nothing, which keeps Karg consistent and affordable, and saves the AI database agents for the cases that need judgment.

Deterministic rules

Handle the straightforward cases: trimming, formats, known codes. They give the same answer on every run and cost almost nothing.

Specialized machine learning

Models built for data preparation infer column types, standardize values, match records, and spot anomalies and likely failures.

AI database agents

Take only the ambiguous few percent. They reason over the column profile and plan, test a fix on a sample, and a separate verifier checks it before it can apply.

  1. Raw Data

    Messy spreadsheets, CSV exports, and database dumps enter the cleaning run.

    Stage Details

    • Typical inputs include survey exports, CRM dumps, vendor CSV files, and manual spreadsheets.
    • Rows can contain inconsistent categories, typos, and mixed formatting patterns.

  2. Profiling Engine

    Karg profiles columns, detects anomalies, and maps likely data types before cleanup.

    Stage Details

    • Detects column intent (state, phone, date, company name) and outlier values.
    • Builds a quality profile to route each field through the right cleaning path.

  3. Rule Engine

    Deterministic rules fix format errors, validate values, and apply fast standardizations.

    Stage Details

    • Applies known mappings and formatting constraints for consistent outputs.
    • Examples: CA -> California, TX -> Texas, ny -> New York.
    • Phone formatting (US): (415)555-1212 -> +1-415-555-1212.
    • Date formatting: 01/15/2024 -> 2024-01-15.
    • Boolean values: Yes, Y, 1 -> true; No, N, 0 -> false.

  4. Specialized ML Models

    A targeted family of machine learning models governing intent, routing, and safety: XGBoost, Random Forest, SVM, Naive Bayes, Isolation Forest, logistic regression, autoencoders, and transformer embeddings (MiniLM).

    Stage Details

    • Identifies column meaning, ranks cleanup options, matches duplicate records, standardizes inconsistent values, detects anomalies across columns, and powers risk assessment for auto-applying changes.
    • Works in sequence to optimize retrieval, measure confidence, and prioritize the review queue.

  5. Quick AI Reasoning Layer

    Fast model reasoning resolves common semantic mismatches that rules alone cannot catch.

    Stage Details

    • Resolves fuzzy variants and abbreviations at high speed for common edge cases.
    • Examples: Calif -> California, Cali -> California.

  6. Advanced AI Reasoning Layer

    Deeper reasoning handles harder edge cases and ambiguous records with confidence scoring.

    Stage Details

    • Handles ambiguous values with stronger context and confidence evaluation.
    • Example: Resolves ‘St. Mary’s’ to ‘St. Mary’s Hospital’ using address context and existing organization names, with confidence scoring and traceable rationale.
    • Example: Resolves ‘AP’ to ‘Accounts Payable’ using finance-column context and existing category values, with confidence scoring and traceable rationale.

  7. Human Review Queue

    Low-confidence records are routed to your team with full context for fast decisions.

    Stage Details

    • Each flagged row includes reason codes, confidence, and suggested normalized value.
    • Teams can approve, override, or add rules that improve future runs.

  8. Clean Dataset

    You export standardized, trustworthy data ready for analysis, dashboards, and models.

    Stage Details

    • Output is normalized, deduplicated, and enriched with confidence + audit metadata.
    • Ready for BI dashboards, analytics workflows, and downstream modeling.

People stay in control

No one acts on a fix they can't check. Karg is built around confidence-based review, not blind automation.

You set the threshold

Every fix has a confidence score. Fixes at or above your threshold (85% by default) apply on their own; the rest wait for a person.

One-keystroke review

The review queue shows each uncertain change with its reason and confidence. A approves, R rejects, E edits.

Nothing hidden

Each change records which layer proposed it and why. Agents never edit your data directly; every fix goes through the same confidence checks.

Karg's Metrics at a Glance

Total records repaired

—

Average Confidence

—

Live totals across Karg runs. Fixes below your confidence threshold (85% by default) wait for review.

Schema InferenceColumn types
Entity ResolutionDuplicate matching
Value StandardizationConsistent formats
Semantic ReasoningContext-aware corrections
Confidence RoutingReview thresholds
Change TraceabilityReason and confidence per change
Audit ExportTransformation log

Watch messy rows transform into clean records

Run a multi-column cleaning example, inspect the detected issues, and trace field-by-field decisions from rules to AI. The records and figures are an illustrative sample, not a customer's data.

Sample dataset (illustrative)

3,482

Rows

12

Columns

217

Issues detected

Sample window below shows 5 representative rows from a larger run.

Detected problems

  • State variations7 variations
  • Company duplicates3 clusters
  • Email typos2 records
  • Phone format inconsistencies4 rows

Column profiling

  • State - Geographic region

    12 unique values, expected domain: US states

  • Phone - Phone number

    5 format variations detected

  • Company - Organization entity

    3 dedupe clusters identified

  • Email - Email address

    2 domain/local-part anomalies

Hover over a row to see its changes explained.

Messy DatasetClean Dataset
NameEmailCompanyStatePhoneNameEmailCompanyStatePhoneDecision
John Smithjsmith@gmail.conACME IncCA(415)555-1212..................
J. Smithjohn.smith@gmail.comAcme IncorporatedCalif4155551212..................
Maria Lopezmaria@globex.coGlobexNwe York212-555-9912..................
Mike Chenmchen@initech.ioInitechTX(713) 555 2281..................
Alicia Gomezagomez@northwind,comNorthwind LimitedCA - USA+1 310 555 7777..................

Explainability

Active row: John Smith

Rule layer handles deterministic normalization and routes unresolved semantic fields to AI review.

Latest transformation event: Name

  • Email

    AI review

    jsmith@gmail.con -> AI review

    Domain typo is detected, but deterministic validation cannot safely infer the canonical mailbox. Confidence: 57%.

    Regex validation failed -> Escalated to AI review queue

  • State

    Rule decision

    CA -> California

    Two-letter abbreviation matched deterministic state dictionary. Confidence: 99%.

    Rule abbreviation match -> Applied canonical state name

  • Phone

    Rule decision

    (415)555-1212 -> +1-415-555-1212

    US phone parser normalized punctuation and added country code. Confidence: 98%.

    Digit extraction -> Formatted to E.164 display

Transformation history

  1. Column profiling confirmed person/company/state/phone schema
  2. Deterministic mappings normalized state and phone
  3. Ambiguous email correction escalated to AI review

Pattern discovery

  • ACME identity cluster detected across rows 1 and 2.
  • Email typo pattern (.con) appears against validated gmail.com frequency.
  • US phone normalization template applied with country code inference.

Cleaning results

0%

Rows standardized

0

Records deduplicated

0

Fields corrected

0

Needs AI review

What you receive

Every run ends with the clean data and the records to check it.

The cleaned file

In the format you uploaded, or as CSV, Excel or JSON. Karg works on a copy and never overwrites what you sent.

A transformation log

One entry for each transformation applied: the column, the action, the reason, the confidence, and whether a rule, a model or AI made the call.

The plan you approved

The cleaning plan and setup you agreed to, as files, so the result traces back to your answers. Save it as a template for the next file.

A run audit report

A check of the run itself: how the file was read, whether every row was preserved, and any change no transformation accounts for.

Technical architecture

Show details

The cleaning engine

A Python worker built on Polars, fed by a job queue, runs each cleanup. Every proposed change passes through the three layers, and a confidence router decides whether to apply it, send it to review, or pass it to the next layer.

Specialized models

Five versioned models handle semantic type inference, value canonicalization, entity resolution, anomaly detection and failure prediction. They are trained in a separate offline pipeline and promoted through quality gates before they serve.

AI database agents

Agents run on open-weight models through zero-data-retention endpoints, so no model provider keeps your data. A fix is tested on a sample and rolled back if the result drifts, and a separate verifier checks it before it is applied.

Isolation and retention

Each organization's data is isolated with row-level security, and files and results are deleted within 72 hours after a job ends. Read the security policy.

Clean data without hiring a data team.

Request platform accessRequest a data assessment