Physical-Context AI.
Intelligence for Retail.

A physical-context AI platform fusing search intent, geo-coordinates, and foot traffic data into store-level retail intelligence. Built and validated with enterprise buyers across multi-national multi-location retail markets.

Status: paused — research line folded back into TNG Shopper’s main pipeline
13
Validated Brands
15
Patent Claims
11
AI Agents
Product Discovery

The Enterprise Problem

Validated through discovery sessions with enterprise retail stakeholders at multi-location retailers operating 15,600+ combined store locations.

verified

Enterprise Discovery Reach

Direct product discovery sessions conducted with retail leadership and decision-makers at Fortune-500 multi-location retailers across grocery, wholesale, pharmacy, and general merchandise.

5,000+
Grocery & General Merchandise Locations
600+
Wholesale Club Locations
10,000+
Pharmacy & Health Retail Locations
Enterprise Retail

13 Buyer Interviews

Validated across Fashion, Sporting Goods, Home & Living, and Jewelry verticals with enterprise retail buyers.

checkroom sprint chair diamond
psychology

Gut-Driven Allocation

Enterprise buyers confirmed that <50% of purchasing decisions are data-driven. The remainder relies on intuition, informal signals, and “gut feeling.”

visibility_off

Blind to Geography

Buyers need demand signals by size, color, and gender—segmented by store location. Current BI systems cannot provide geo-specific SKU-level intelligence.

block

No Unified Signal Layer

Search intent, visitor geo-coordinates, and product catalog data exist in separate silos. No system connects these signals into a unified, store-level demand view—leaving buyers without spatial evidence.

System Design

How It Works

A conversational multi-agent system. Ask a question in natural language—11 specialized agents collaborate to deliver store-level intelligence.

Conversational Interface

A single orchestrator receives natural language queries in Hebrew or English, classifies intent, and dispatches work to the right combination of expert agents—no manual configuration required.

  • route
    OrchestratorIntent classification and dispatch. Routes each query to the optimal agent chain based on context.
  • smart_toy
    5 Data AgentsGeo-matching, visitor-search fusion, behavior aggregation, spatial visualization, and trend analysis.
  • task
    5 Business AgentsOffline advertising, store audiences, upselling opportunities, top searched categories, and emerging trends.
Query → Intelligence → Action
“Which sneakers are trending near Haifa?”
Orchestrator — Intent → Geo + Trend agents
GeoMatcher TrendAnalyzer SearchCombiner
Structured demand report → Looker Studio
database

Three Signal Sources

  • Product Catalog
    SKU attributes + store geo-coordinates
  • Visitor Location
    Pseudonymous foot traffic · Hourly
  • Search Intent
    GSC queries, impressions, CTR · Daily
verified

Reliability Guardrails

  • ≥50 searches/month
    Minimum signal threshold per product
  • ≥3 months stable data
    Temporal consistency before inference
  • ≥80% label completeness
    Data quality gate before any output
science

Model Selection

  • PaLM 2 Bison-002Production
  • PaLM 2 Bison-001Stable baseline
  • Gemini 1.0 ProEvaluated · Not selected

Bison-002 selected for consistent Hebrew geo-query handling and structured output reliability.

memory

Built on Vertex AI

Vertex AI Agent Builder for orchestration. Vertex AI RAG and Vector Search for retrieval-augmented generation. Multiple tool executions per query chain. All agents operate within Google Cloud’s enterprise security boundary.

Compliance

Privacy & Security

Enterprise-grade compliance by design.

Anonymization

Mathematical anonymity. Strict k-anonymity (≥10) protocols ensure zero PII ever enters the reasoning layer.

Data Classification

Vertical-specific ontologies. Tailored classification boundaries for Electronics, Fashion, Beauty, Food, Home, and Pet sectors.

Access Control

3-tier isolation (Public / Licensed / Enterprise) backed by strict cloud IAM boundaries and zero-trust network policies.

“Crowd Memory” Architecture

No persistent user profiles. Visitor identifiers are ephemeral and cryptographically hashed. All predictions are derived from aggregate cohort behavior within H3 micro-regions, not individual user histories. The system explicitly avoids persistent user profiling—using “Crowd Memory” instead of “User Memory.”

Intellectual Property

Innovation

3 independent + 12 dependent patent-pending claims.

CCR = Purchases / (Temporally Relevant + Geographically Proximate Impressions)

Unlike standard conversion rate, CCR measures only impressions that are both temporally relevant and geographically proximate—the efficiency of physical-context matching.

Independent Claims (Patent Scope)

Claim Scope
1. Method Claim End-to-end process: ingest physical-context signals → transform via sub-agents → compute Demand Index → trigger actions → generate structured output → publish → feedback loop.
2. System Claim Physical architecture comprising 7 subsystems (ingestion, transformation, computation, trigger, generation, publication, feedback) with closed-loop learning.
3. Medium Claim Computer-readable medium storing instructions that, when executed, perform the method of Claim 1.

Key Innovations (Dependent Claims)

Innovation Description
H3 Geospatial Indexing O(1) spatial lookups mapping consumer intent to precise retail coordinates via Uber’s H3 hexagonal grid.
Grammar-Constrained Decoding Enforcing strict structural compliance on LLM outputs via context-free grammars to prevent supply chain execution errors.
Z-Score Trigger Gating Statistical thresholding that halts redundant inference when signals fall within normal deviation, reducing compute overhead.
Variant-Level Intent (VSIS) Isolating demand signals down to the specific product variant (size, color, gender) within a geographic radius.
Zero-Result Signal Analysis (ZRFAS) Inverting failed retail searches into predictive signals for unmet local demand.
Demand Index (DI) Composite score combining search velocity, geo-concentration, and temporal signals into a single actionable metric per SKU per location.

Implementation

Vertex AI Agent Builder in action.

Enterprise Validation

What Buyers Told Us

13 enterprise interviews validated both the market gap and the barriers to adoption.

4

retail verticals confirmed the gap: Fashion, Sporting Goods, Home & Living, and Jewelry.

3–5

stakeholders per approval chain—from buyer to CEO to VP Commerce. Enterprise adoption friction is structural, not product-related.

Pilot-Ready

Multiple respondents shared contact details for follow-up pilots—a direct product-market fit signal from enterprise buyers.

lightbulb

Validation Outcome

Enterprise buyers across all four verticals confirmed strong demand for store-level, variant-specific intelligence—a capability none of their existing BI systems provide. However, the 3–5 stakeholder approval chain per retailer extends sales cycles well beyond early-stage runway constraints. This structural friction—not product skepticism—directly informed the strategic pivot to Autonomous Pipeline.

The Decision

Strategic pause driven by data.

Product-Market Fit

Validated. High demand confirmed through 13 enterprise buyer interviews across multi-national multi-location retail markets. Buyers confirmed <50% of decisions are data-driven—a clear market gap.

Big Data Dependency

Reliable spatial intelligence requires large-scale, longitudinal datasets aggregated across multiple regions and retail sectors. Without sufficient client adoption—and the data breadth it generates—outputs risk distribution bias, concept drift, knowledge cutoffs, and regional skew. Elysium’s value scales with adoption, not in isolation.

Stakeholder Friction

High. Enterprise validation confirmed 3–5 stakeholder approval chains (Buyer, CEO, VP Commerce)—extending sales cycles beyond runway constraints.

Outcome

Doubled down on autonomous content generation (Autonomous Pipeline)—a frictionless adoption layer that scales the client base and the data surface simultaneously. Once adoption reaches critical mass, Elysium activates as the next-phase intelligence product: an upsell that transforms the Pipeline’s aggregated data into store-level demand decisions.

Addressable Market: Retail Location Intelligence ($2.8B → $10.9B by 2033) · AdTech ($217B) · MarTech ($6B, 30%+ CAGR) · Loyalty Systems ($12.95B, 18% CAGR)

The entire lifecycle (discovery → research → development → testing → evaluation → decision) was led end-to-end by Daniel Manzela.

Published Research

Thought leadership in physical-context AI.