Grocery Store Digital Twin for Wayfinding and Guest Services
Shoppers could not find products, staff spent their shift answering the same location questions, and the store had no live spatial picture of aisles, displays, or queues. Store layouts change weekly, so any static map or app directory is wrong within days.
- Live twin
- Store Model
- Jetson + Dell GB10
- Edge Compute
- None
- Video Egress
The operational problem
Grocery is one of the hardest retail environments to model. Aisles are reorganized, end caps rotate, seasonal displays appear overnight, and out-of-stocks move products to different shelves. A wayfinding experience built on a static planogram is wrong almost immediately, and guest-service staff absorb the gap. Answering "where is the tahini" correctly requires a spatial model of the store that updates itself, and that model has to be built and queried inside the store — shoppers are on camera, and that footage cannot be shipped to a cloud analytics service.
What Enfuse built
- Camera-based perception in-aisle for shelf, fixture, and display recognition
- A spatial store model — aisles, bays, shelves, end caps — kept current from live observation
- Product-to-location mapping reconciled against the retailer's item master and planogram feed
- A guest-service assistant that answers location, availability, and substitution questions against the twin
- Kiosk and associate-handheld interfaces with turn-by-turn in-store directions
- Queue, dwell, and congestion signals surfaced to store leadership in real time
- Anonymization at capture — no faces, no identity, no biometric matching
Architecture
- See — in-aisle cameras stream to NVIDIA Jetson devices for local inference
- Understand — shelf, fixture, and display recognition at the edge on Jetson
- Locate — observations resolved into the store's spatial coordinate model
- Reason — the Dell GB10 on site hosts the twin and the guest-service assistant
- Act — wayfinding at kiosks and handhelds; congestion and gap alerts to staff
Hardware and software
- NVIDIA Jetson Orin (in-aisle edge inference)
- Dell GB10 on-site workstation (digital twin + assistant inference)
- Dell networking and store infrastructure
- TensorRT-optimized vision models, multi-object tracking, spatial indexing
- Integration with item master, planogram, and inventory systems
How Claude Enterprise was used
Claude Enterprise ran alongside the in-store system as the engineering and operations layer: Claude Fable 5.1 built and tested the operational twin, and Claude Enterprise helped the team maintain it, and helped merchandising staff keep store data accurate.
Claude Fable 5.1 builds the twin
Enfuse used Claude Fable 5.1, Anthropic's most capable model for long-running engineering work, in Claude Code to build the operational twin itself: the store's spatial coordinate model, the aisle-bay-shelf hierarchy, and the logic that turns Jetson detections into live shelf and fixture state.
Days-long tasks, not prompts
Fable sustains multi-day, asynchronous work. It was handed whole milestones — ingest a new store layout, generate its twin, wire it to the planogram feed, and prove it against recorded test runs — and checked in with progress updates instead of needing step-by-step direction.
The whole store in one context
With a 1M-token context window, Fable reasoned over a complete store at once: CAD floor plans, fixture lists, the full item master, planogram exports, and the twin's own code, catching mismatches that piecemeal review would miss.
Simulation before the store floor
Fable generated synthetic layout changes — a moved end cap, a seasonal display, an out-of-stock shuffle — and test scenarios against them, so twin updates and wayfinding answers were proven in simulation before reaching a live store.
Rolling out store after store
Each new store became a repeatable job: Fable adapted the twin template to that store's layout, produced the integration checklist and acceptance report, and flagged anything that needed an engineer's judgment.
Zero data retention
Fable 5.1 ran under zero data retention, so design files and store data sent to the model were not stored by Anthropic — and shopper video never entered a session at all.
Claude Code with parallel agents
Engineers ran several Claude Code agents at once from a single task board — one tuning TensorRT perception models, one writing planogram reconciliation, one building kiosk screens — each on its own branch, with sub-agents for tests and code review before a human approved the merge.
Hooks and repository guardrails
Claude Code hooks enforced the project's rules automatically: linting and tests on every change, blocked writes to camera-ingest and anonymization code without sign-off, and no secrets or store footage ever pulled into a session.
Cowork for long-running data work
Merchandising staff handed Claude whole jobs — reconcile this week's item master against the planogram feed and the twin's shelf observations — and got back a finished exception report. The work kept running in the cloud after they closed the laptop.
Excel to PowerPoint in one pass
Claude carried context from the gap-analysis spreadsheet straight into the weekly store-review deck, so the analysis and the slides that present it came out of the same session.
Custom plugins per team
Enfuse packaged skills, connectors, and sub-agents into plugins for each role — a reset-planning plugin for merchandising, a runbook plugin for store IT, a review plugin for engineering — published through the retailer's private plugin marketplace.
Skills that encode store standards
Store layout rules, signage standards, and escalation steps were written once as skills, so every answer and every document followed the retailer's way of working without re-explaining it.
Connectors and MCP servers
Approved connectors linked Claude to GitHub, Slack, and Google Drive. A custom MCP server exposed read-only, anonymized twin summaries — aisle changes, gap counts, congestion trends — never raw video or live camera access.
Projects with a large context window
Integration specs, item-master exports, and full store layouts sat in shared Projects, so engineers and operations reasoned over the complete picture in one session instead of in fragments.
Custom roles, SSO, and SCIM
Access came from the retailer's identity provider. Custom roles decided who could use chat, Cowork, Claude Code, web search, and each connector, and seats were added or removed automatically as people joined or left teams.
Audit logs, Compliance API, and data controls
Every session was logged and streamed to the retailer's security tools. Group spend limits kept usage predictable, and enterprise data controls meant prompts and files were never used for model training and were kept only as long as policy allowed.
The boundary did not move: shopper video and live inference stayed on Jetson and the Dell GB10 in the store. Claude Enterprise worked only with code, documents, and anonymized summaries.
Deployment environment
On-premises in-store. Camera inference on Jetson, twin and assistant on the Dell GB10 in the back office. No video egress.
Results
- Wayfinding answers grounded in the store's current layout rather than a static directory
- Guest-service load shifted from associates to self-serve kiosk and handheld interactions
- Store leadership gets live congestion and shelf-gap signals from the same perception layer
Client engagement. Store-level figures are withheld at the customer's request; the video shows the deployed digital twin and wayfinding experience.
Discuss this pattern against your environment
Bring your constraints — sensors, security boundary, latency budget — and we will tell you what is realistic.