NEXWAWEINFOTECH
All projects

OCR + decision intelligence

Warehouse OCR label reconstruction and inventory matching engine

A traceable Python engine that rebuilds fragmented OCR output, corrects common misreads, searches inventory for candidates, scores ambiguity, and flags items photographed in the wrong bin.

Platform foundation

PythonOCRLevenshteinData pipelinesScoring heuristicsCSV reporting

The challenge

Raw OCR output was fragmented, duplicated, incomplete, and frequently imperfect. Exact string matching could not reliably decide which detected label belonged to which inventory record, and a confident wrong answer was more damaging than no answer at all.

The solution

We built a staged matching engine that reconstructs labels from their spatial fragments, applies correction rules, then scores every plausible inventory candidate. When two candidates score too closely the engine refuses to guess and routes the case to a person.

End-to-end workflow

How the product moves work forward

  1. 01

    Collect detections

    Take raw OCR fragments with their bounding boxes and confidence values.

  2. 02

    Reconstruct labels

    Group fragments spatially and rebuild the intended label text.

  3. 03

    Search candidates

    Run exact, prefix, suffix, numeric-prefix, and Levenshtein fuzzy searches against inventory.

  4. 04

    Score and decide

    Combine multiple signals, apply ambiguity protection, and either match or escalate.

  5. 05

    Publish evidence

    Emit CSV summaries and per-image evidence packs supporting every decision.

Detailed capability map

What was designed into the platform

Label reconstruction
  • Spatial OCR fragment grouping
  • Normalization and correction rules
  • Duplicate and partial fragment handling
  • Character confusion recovery
  • Confidence propagation
Candidate search
  • Exact match
  • Prefix and suffix match
  • Numeric-prefix match
  • Levenshtein fuzzy matching
  • Inventory index lookups
Decision intelligence
  • Multi-signal candidate scoring
  • Ambiguity protection thresholds
  • Explicit no-match outcomes
  • Photographed-location comparison
  • Bin mismatch flags
Traceability
  • CSV run summaries
  • Per-image evidence packs
  • Score breakdown per candidate
  • Reproducible runs
  • Reviewable escalation queue

Roles and permissions

The right workspace for every participant

Permissions support the real operating model instead of giving every user the same controls.

Inventory analyst

Reviews escalated ambiguous matches and confirms the correct inventory record.

Warehouse supervisor

Acts on bin mismatch flags where an item was photographed in the wrong location.

Data engineer

Tunes correction rules, scoring weights, and ambiguity thresholds against real runs.

Operational value

  • Turns unusable OCR fragments into matchable label text
  • Avoids confident wrong matches through ambiguity protection
  • Flags items photographed in the wrong bin
  • Leaves an evidence trail behind every automated decision

Growth opportunities

  • Learned scoring from analyst corrections
  • Client-specific label grammars
  • Live review interface for escalations
  • Accuracy dashboards by site and camera

Build on this experience

Have a workflow that needs its own product?

Tell us what your team does manually today. We will help identify the most useful first version.

Start a project See other projects