Product Data Quality Agent
From messy product documents to reconciled product data
A working agent that extracts product data from invoices and customs documents, reconciles records across sources, handles known cases with rules, and sends uncertain decisions for human review.
01 — The business problem
Product information often arrives across invoices, supplier files and customs documents. The same product can appear with different identifiers, OCR errors, descriptions and commercial values. Someone then has to work out what belongs together, which differences are harmless, which records can be corrected safely, and which cases need judgment. The problem becomes larger when the same process repeats across hundreds or thousands of documents.
02 — The workflow
- Documents are uploaded and their contents are extracted with the original source evidence preserved.
- Explicit rules handle known cases such as safe identifier normalization, malformed values and duplicate documents.
- Evidence across documents is used to establish product identity where it can be supported.
- AI is used only for cases that require interpretation, such as ambiguous product matching.
- AI suggestions are validated and either proposed for change or sent for review. They never write directly into the canonical product data.
- The user reviews unresolved cases and exports the product master, commercial evidence, issues and audit information.
03 — What I built
- A Python processing pipeline for document extraction, data-quality checks and product reconciliation.
- A bounded AI layer for ambiguous cases, with validation before a suggestion can enter the review workflow.
- A human-review system with approve, reject and defer decisions.
- Provenance and audit tracking so extracted facts and system decisions can be traced back to their source.
- A web application covering the full flow from upload through processing, product data, review and export.
- A reconciled Excel workbook plus raw outputs for further use.
04 — Tools used
Python · FastAPI · OpenAI API / GPT-5.6 Luna · Lovable · Render · GitHub
05 — Commercial use case
A company importing, distributing or managing products can use the system to turn recurring document-based reconciliation work into a controlled workflow. The same system pattern can be adapted to other processes where teams repeatedly extract information, apply known rules, investigate ambiguous cases and review exceptions.
06 — Evaluation
Representative evaluation run
- 8
- 273
- 26
- 100
- 29
- 89
These numbers describe one evaluation dataset, not general product metrics. The system was tested against representative document structures, failure cases and deliberately ambiguous evidence.
Current limits
The current version supports selected document layouts. Some commercial evidence remains unresolved, and AI-generated proposals still require human review.