Product Data Quality Agent

From documents to reconciled product data

Upload product, supplier and customs documents. The system extracts the available data, checks it, connects records that refer to the same products, and sends uncertain cases for review.

Document extraction · rules-based checks · product reconciliation · AI-assisted review

The problem

Product data is often spread across documents

  • The same product may appear across invoices, supplier files and customs documents.
  • Names and identifiers can differ.
  • OCR can introduce errors.
  • Prices, quantities and other values may appear in different places.
  • Someone then has to compare the documents and decide what belongs together.

The source data contains inconsistent identifiers, OCR errors, fragmented observations and incomplete product context.

What the system does

Handle the clear cases automatically. Review the uncertain ones.

The system extracts what the documents say, checks known conditions, connects evidence across documents and separates decisions that can be made automatically from those that need review.

Deterministic rules run first. AI is used only where interpretation is needed. Proposed changes are not applied automatically.

How it works

From documents to reviewable data

  1. 1

    Upload

    Add supported invoices and customs documents.

    PDF input with source registration and provenance tracking.

  2. 2

    Extract

    Read the available product, quantity, price and document information.

    Each extracted observation keeps its source document, location and original text.

  3. 3

    Check

    Detect known errors and safe normalizations.

    Rules handle cases such as approved identifier normalization, invalid values and duplicate sources.

  4. 4

    Reconcile

    Connect records that refer to the same product.

    Product identity is established only when the available evidence is strong enough.

  5. 5

    Interpret uncertain cases

    Use AI where the rules cannot safely decide.

    Semantic matching is validated before it can become a proposed change.

  6. 6

    Review

    Approve, reject or defer decisions that still need a person.

    Review decisions are recorded separately and do not silently alter canonical data.

  7. 7

    Export

    Download the reconciled data, evidence and review information.

    Outputs stay traceable back to the documents they came from.

Decision model

Rules first. AI where needed. Human review when uncertain.

  • Known cases are handled with explicit rules.
  • AI is used for cases that require interpretation.
  • If the evidence is still not sufficient, the system leaves the decision for a person.
AUTO_FIXPROPOSE_CHANGENEEDS_REVIEWNO_CHANGE

Semantic suggestions never write directly to the canonical product master.

What the user gets

Structured data with the evidence behind it

The output is not just a cleaned file. It keeps the source evidence, unresolved cases and the reason behind each decision.

  • Product master
  • Extracted observations
  • Commercial evidence
  • Proposed changes
  • Review queue
  • Issues
  • Audit trail
  • Reconciled workbook

The workbook is derived from the canonical outputs and preserves unresolved commercial evidence instead of inventing links.

Example run

One evaluation run

8
input documents
273
extracted observations
26
canonical products
100
automatic normalizations
29
proposed changes
89
items requiring review

These numbers describe one evaluation dataset. They are not general product metrics.

The evaluation used real document structures with sanitized public demo data.

Why human review remains

The system does not turn uncertainty into fact

  • If a value is unclear, the system keeps it unclear.
  • It does not invent missing values.
  • It does not silently repair damaged identifiers.
  • It does not automatically merge products just because they look similar.
  • It does not automatically apply AI-generated changes.

Provenance, confidence, decision source and audit information are retained for review.

Case study

The main work was defining the right workflow

This project started with a specific data problem.

The work was to understand what the documents contained, which decisions could be made safely, where the evidence was weak, and what output would actually be useful.

The resulting system combines normal software, explicit rules, AI and human review.

The architecture separates extraction, reconciliation, deterministic checks, semantic reasoning and review instead of treating the LLM as the whole system.

Architecture

How the system is structured

  1. Documents
  2. Structured data
  3. Checks
  4. Product reconciliation
  5. Review
  6. Export

Web application → FastAPI → Python processing pipeline → deterministic rules → OpenAI semantic layer → review workflow → canonical outputs

Current limits

Limits of this version

  • This version supports selected invoice and customs-document layouts.
  • Some document layouts still require new extraction rules.
  • Some commercial evidence cannot yet be linked to a specific product.
  • AI decisions can vary between runs.
  • Human review is still required before proposed changes are applied.

The current MVP was validated against a representative document set rather than designed as a universal document processor.

Relevance to other workflows

The same method can be applied to other workflows

This exact agent is specific to product-data reconciliation.

The broader method applies wherever a team repeatedly reads information, checks rules, reconciles records, handles exceptions and makes similar decisions.

The data model, rules, integrations and review process would change for each use case.

The reusable part is the system pattern: deterministic processing + bounded AI reasoning + validation + human review.

Have a similar workflow?

If your team repeatedly reviews documents, reconciles data, checks exceptions or makes the same decisions across many records, send me a short description of the process.

LinkedIn