US Utility Patent #19/562,766 — at examiner

The shelf can finally tell you why.

A new class of retail intelligence — semantic, real-time, planogram-bound. No raw audio. No transcripts. No shopper identification. Only meaning, only what matters.

A privacy-by-architecture system designed in London · Florida · Kyiv
shelf mic
VAD gate · 1.5s window
MFCC → INT8 classifier
planogram-locked taxonomy
k-threshold release
signed aggregate · MQTT
entirely on-device · buffer zeroized · only counts leave the edge
Product

A small device, a precise discipline, an entirely new measurement.

A ShelvesLab unit sits quietly above a shelf segment. It listens for the brief acoustic moment of a shopper's spoken reaction — and it translates that moment, on the device itself, into a single bounded semantic count. Nothing more leaves the edge.

listening
Form factor
A compact shelf-edge unit, fixed above the planogram. USB-C or PoE.
Sensing
Near-field acoustic capture only. No camera. No facial data. No identifiers.
Inference
On-device. INT8 quantized classifier mapped to a planogram-bound taxonomy.
Retention
Zero. The buffer is overwritten in volatile memory after every inference.
Release
k-thresholded counts only. Sparse buckets are dropped before transmission.
Integration
MQTT over TLS, ed25519-signed payloads, multi-tenant isolation at ingest.
Reference hardware: ESP32-S3 class. Bill-of-materials under USD 60.
The problem

Retail measures conversion. It cannot measure rejection.

A shopper picks up the bottle, frowns, says "too sweet," puts it back. That micro-event is the single most expensive informational loss in physical retail — and no instrument captures it.

lagging

Point of sale

Shows what converted. Silent on what was considered and refused.

delayed

Surveys & reviews

Post-hoc, biased by memory, only from those who chose to speak.

invasive

In-store cameras

Track dwell and gaze. Cannot capture meaning. Compliance risk: GDPR · BIPA.

"The store does not lead you. It places you in the correct position — and the choice you make after that feels entirely your own."

— Ievgeniia Shneider, FCIM · LIGA.net interview, "The architecture of persuasion"

What is new

A new category of measurement — defined, named, and protected.

ShelvesLab does not optimize an existing metric. It introduces the shelf signal — a bounded, context-linked, semantically interpretable indicator of consumer meaning that arises during product consideration, and that no prior instrument captured.

01

A field, not a feature.

Defined across a 10-paper open research series and a Library-of-Congress book. The category itself is new — semantic shelf intelligence. ShelvesLab is its first commercial instrument.

02

A technique, not a model.

Planogram-locked voice-to-tag inference: the device's interpretive vocabulary is hardware-constrained by the shelf segment it serves. No general-purpose language model. No speech-to-text path.

03

A property, not a policy.

Privacy enforced in firmware: the buffer is zeroized after every inference. The device cannot retain audio because it has no code path that could. Restraint is structural.

Protected by US Utility Patent #19/562,766 — at examiner
Advantage

Built to be defensible — technically, legally, and commercially.

Approach
Captures reason
Real-time
Compliance-safe
Scales economically
ShelvesLab
POS analytics
Surveys / reviews
Computer vision (Trax / Quividi)
Cloud-based speech analytics

Patent protected

The combination of planogram-locked taxonomy + voice-to-tag + k-threshold release is filed as US Utility Application #19/562,766.

Compliance defensible

No facial data. No transcripts. No identifiers. Built for GDPR, CCPA, and Illinois BIPA from the firmware up — not from the privacy policy down.

Economically scalable

Edge inference. No continuous audio streaming. No cloud GPU bill. Sub-USD-60 BOM per shelf segment.

The economics

Premium retail spends billions guessing. We make it measurable.

In the categories where ShelvesLab serves first — fragrance, beauty, wine — a single mis-placed product can cost a brand more than a year of shelf rent. The cost of getting it wrong is enormous. The cost of knowing within an hour is a recurring subscription.

≈ 70%
of fragrance sampling ends in rejection

High-consideration beauty has the highest informational loss in retail. Every refusal is currently invisible to the brand.

≈ $9B
global retail analytics market

Currently dominated by sales-data tools and computer-vision vendors. Voice-to-tag is a vacant lane.

< $60
bill-of-materials per shelf segment

ESP32-S3 class hardware. No cloud GPU. Recurring revenue against negligible per-shelf cost.

Sold to

Brand intelligence teams

FMCG and beauty houses that need to understand the qualitative reason behind shelf rejection — before the return arrives, before the SKU rotates, before the season ends.

Priced as

Per-shelf subscription

A monthly recurring fee per active planogram-bound device. Hardware included. Dashboard included. Software updates cryptographically signed.

Defended by

A US utility patent

The architecture is patent-protected at the examiner stage. PCT international filing window open until February 2027. Licenseable to retailers and brands at scale.

Technology

Five stages.
All on-device. Only counts leave the edge.

Each device runs the same five-stage pipeline. The privacy boundary is enforced in hardware before any byte is transmitted.

01

Audio capture

A voice-activity detector opens a short capture window (0.25 – 2.0 s) only on speech-like energy. Samples land in volatile RAM. The buffer is overwritten — every byte set to zero — immediately after the next stage finishes. No NVS, no flash, no SD. There is no code path that can persist raw audio.

VAD gate · ephemeral RAM buffer
02

Feature extraction

The buffer is converted into compact MFCC features — the same compact spectral summary used in every modern speech system. MFCC is lossy by construction; raw audio cannot be reconstructed from it under any practical adversarial model.

MFCC · log-mel · fixed-point
03

Planogram-locked classifier

The device is provisioned with a planogram_id. That identifier deterministically loads a narrow acoustic model and a constrained taxonomy. A fragrance shelf can only emit fragrance tags. A wine shelf can only emit wine tags. There is no general-purpose language model on the device and no path to one.

INT8 TFLite · TinyML
04

Privacy engine

Tags accumulate inside fixed time buckets, with sub-window deduplication. A bucket is released only if at least one tag reaches the k-threshold. Sparse buckets are dropped at the device — nothing crosses the network.

dedupe · bucket · k-threshold
05

Secure ingest

Released aggregates are signed with the device's per-unit key, transmitted over MQTT-over-TLS, schema-validated at the broker edge, and persisted into a tenant-isolated storage layer. Brand A never sees Brand B's aggregates — even on the same shelf bay.

MQTT · TLS · ed25519 signed
What actually leaves the device

A telemetry payload, in full.

Notice what is absent by design.

telemetry payload — v0
{
  "schema_version":    "v0",
  "device_id":         "korvo-007",
  "planogram_id":      "sephora.nyc.5th-ave.shelf-007",
  "planogram_version": "2026.05.01",
  "taxonomy_version":  "fragrance.v1",
  "firmware_version":  "0.4.2",
  "bucket_start":      "2026-05-13T14:00:00Z",
  "bucket_end":        "2026-05-13T14:10:00Z",
  "k_threshold":       3,
  "tags": [
    { "tag_id": "SWEETNESS_NEG",    "count": 7, "avg_confidence": 0.81 },
    { "tag_id": "PRICE_HESITATION", "count": 5, "avg_confidence": 0.79 },
    { "tag_id": "GIFT_INTENT",      "count": 3, "avg_confidence": 0.74 }
  ],
  "device_sig": "ed25519:..."
}
No per-event timestamps. No audio. No transcripts. No demographics. No identifiers. No retailer can subpoena what was never collected.
Live demo

Real telemetry.
From four simulated shelves, right now.

Every tag below originated as a fleeting acoustic event at a virtual retail shelf, processed through the same pipeline a production device runs. The privacy engine releases only what passed the k-threshold. Nothing else exists.

dashboard.shelveslab.com Open in new tab

Tip — the live feed updates as the simulator releases new k-thresholded buckets every few seconds.

Ievgeniia Shneider, FCIM — founder of ShelvesLab
Ievgeniia Shneider, FCIM
Founder & Inventor · ShelvesLab
ORCID 0009-0008-5425-5283
The founder

A decade in niche fragrance.
A single quiet patent.

Ievgeniia Shneider spent ten years inside the most discriminating corner of the beauty industry — niche perfumery — before she became the principal researcher and inventor behind ShelvesLab. She holds the Fellow grade of the Chartered Institute of Marketing, the British professional body operating under a Royal Charter granted by Queen Elizabeth II in 1989. The grade is held by fewer than two percent of the institute's thirty thousand members.

Her book on the architecture of shelf intelligence is catalogued in the U.S. Library of Congress. Her ten-paper open research programme on semantic shelf intelligence runs in parallel with her commercial work. The utility-patent application for the architecture this site describes — US 19/562,766 — is currently at examiner.

She works between Florida, London and Paris. She does not appear in industry rankings, and she does not seek to.

"In a world where everyone speaks about themselves, I chose the world that speaks for you."
— Vogue Ukraine cover story, "The woman who rewrites the rules"
Professional
FCIM

Fellow, Chartered Institute of Marketing — Royal Charter, 1989

Bibliographic
Library of Congress

Author of "The Shneider Framework for Semantic Shelf Intelligence"

Research
10-paper series

ELITE WORKS — open scholarship on shelf intelligence and privacy-preserving consumer analytics

Invention
US #19/562,766

Utility application at examiner · priority February 2026

"Silence is not the state before recognition. Silence is an instrument."

— Vogue Ukraine, "The woman with the King's seal"

Contact

Three first-wave pilots.
One global brand. One niche house. One regional retailer.

Devices, training, and dashboards are provided for the pilot wave. You bring the planogram. Write to Ievgeniia directly.

Or write directly: s.ievgeniia@shelveslab.com