Client systemRetail intelligence · Computer vision

Shelfr

Shelf intelligence under real retail conditions.

Partner / contextShelfr
Binding constraint

Recognize near-identical products reliably enough for the shelf decision downstream.

Binding constraintRecognize near-identical products reliably enough for the shelf decision downstream.
01Field image
02SKU recognition
03Shelf structure
04Compliance + availability
05Store action

The vision system behind a shelf platform used by global brands, judged by what a mistake costs the store rather than by an average score.

5 yearsAI architecture and lifecycle engineeringFounder-led record
4 boundariesVariant · occlusion · lighting · catalogue change
Release gateConfusion evaluated by product consequence
EvolvingArchitecture preserved through model change

Retail shelves compress difficult perception problems into one frame: near-identical packaging, partial visibility, reflective surfaces, inconsistent lighting, changing planograms, and catalogues that do not stay still.

The useful output was not a label in isolation. Recognition had to support shelf structure, compliance, availability, and a store action. An error at the SKU boundary could therefore become an operating error downstream.

Nixense led the AI architecture across the path from field imagery to product-level shelf understanding: partitioning the classification space, making catalogue change explicit, exposing costly confusions, and preserving the operating contract through new model lineages.

Evidence boundary

Shelfr was the product partner and award recipient. This record does not present P&G as a Nixense client or attribute Shelfr’s later Agentic AI feature to Nixense.

  1. 01Partition visually adjacent variants so errors remain diagnosable.
  2. 02Make catalogue change explicit instead of treating the label universe as fixed.
  3. 03Separate perception from shelf logic so each failure can be located.
  4. 04Evaluate field variation rather than clean reference imagery alone.

Evaluation joined ground truth, predicted variant, and shelf consequence. A confusion that changes compliance or availability can block release even when an aggregate benchmark looks strong.

Evaluation became a release gate and diagnostic system — not a benchmark reported once and forgotten.

The durable asset was the operating contract around the model: declared inputs and outputs, catalogue behavior, evaluation lineages, and the software boundary consuming recognition. That made incremental migration and re-evaluation possible.

Shelfr’s published platform figures describe more than 200 global brands and average improvements of 25% in shelf compliance and 22% in on-shelf availability. These are Shelfr’s published platform figures — not measurements independently re-derived by Nixense.

P&G named Shelfr among 40 recipients in its 2025 External Business Partner Excellence Awards. Shelfr received the recognition; Nixense’s claim is narrower: its founder led the architecture and model lifecycle of Shelfr’s perception system across the preceding five years.

Record close
The model was never the whole shelf-intelligence system. Dependability came from how the catalogue was partitioned, how data entered each lineage, how the consequential error was measured, and how the architecture evolved without losing its operating history.

Working on a problem like this?

Tell us what the system needs to do, what exists today, and where you need help.

Discuss a system
Principal review[email protected]Pakistan · Working internationally