AutRi captures shelves; GMetri reads them against the planogram; the Agentic Command Centre routes the fix to PulsePoint. The next AutRi scan closes the ticket.
95%+
measured at FreshPik Powai pilot
98%
per-SKU bounding boxes
20–30K sqft
single-store sweep
< 24 hr
next scan confirms the fix held
Phase 1 with Fynd Nucleus is running at FreshPik Powai. AutRi executes scheduled scans, draws bounding boxes around recognised SKUs, and produces brand-occupancy and facing heatmaps. A human still has to log in, interpret the heatmap, and tell someone to fix what's off. The agentic flow removes the human from the middle of that loop.
Today · FreshPik Powai
Scheduled scan sessions captured per zone and fixture · 20–30K sqft per run
Internal CV models draw SKU bounding boxes against AutRi's own SKU master · 98% recognition
Output: brand-by-shelf-level + facing-count heatmaps in AutRi's dashboard
Human loop: a Cluster Manager logs in, interprets the heatmap, manually raises the fix request
In the agentic flow
No more human interpretation — tagged shelf captures flow directly into GMetri via API
Context-aware scanning — schedule dynamically tied to restocking windows and the promotion calendar
The closer of the loop — a PulsePoint ticket marked “Done” stays open until AutRi's next scan of that fixture confirms compliance
Trust comes from verification, not from the staff member's word
~t.compress(q:75)~t.toFormat(f:webp)/fynd-web/solutions/autri/autri-01-sessions-dashboard-v2.jpg)
Sessions surface — scheduled scans across the store, with per-session status.
~t.compress(q:75)~t.toFormat(f:webp)/fynd-web/solutions/autri/autri-02-session-detail-v2.jpg)
Session-detail surface — 100 scan tasks broken down by 50 grocery categories.
~t.compress(q:75)~t.toFormat(f:webp)/fynd-web/solutions/autri/autri-05-brand-shelf-analytics.jpg)
Analytics output — brand-by-shelf-level + facing-count heatmap. Useful, but the human still has to read it and decide what to do.
Planogram compliance means the live shelf matches the plan in every store, all day, every day, across six dimensions. Each one is what the AutRi scan reads, what GMetri interprets, and what the Agentic Command Centre decides on.
Is the SKU on the shelf at all. The most basic compliance question and the one that hits revenue first.
Non-compliance example: empty slot for a top-velocity SKU.
Is the right SKU on the right fixture and shelf level. Wrong brand on the premium shelf is wasted shelf-space and a confused customer.
Non-compliance example: wrong brand of cooking oil on the premium shelf.
Are the expected number of facings present. Under-faced promotional SKUs lose visibility and miss the demand spike.
Non-compliance example: 2 facings instead of the planned 4 for a promotional SKU.
Does the price tag match the current price master. Stale tags during a promo break customer trust at the till.
Non-compliance example: old price tag on a live promotional SKU.
Are promotional end caps and POS displays set up. Promotions drive a disproportionate share of grocery sales; a missing end cap is direct revenue loss.
Non-compliance example: missing end cap for the week's featured SKU.
Is the shelf set up to the visual standard. Toppled, stacked, or damaged products signal a store that is not running its routine.
Non-compliance example: products toppled, stacked incorrectly, or damaged.
~t.compress(q:75)~t.toFormat(f:webp)/fynd-web/solutions/autri/autri-03-shelf-bounding-boxes.jpg)
Per-SKU detection on a high-density grocery wall. Every box is a candidate for one of the six checks above.
~t.compress(q:75)~t.toFormat(f:webp)/fynd-web/solutions/autri/autri-04-shelf-bounding-boxes-2.jpg)
Per-brand attribution on a cooking-sauces fixture. Brand × shelf level × facing count are computed off the same capture.
Every layer exists to serve the planogram-compliance use case. No more, no less. The Data layer sees the shelf, the Intelligence layer judges it, the Action layer fixes it, and the Data layer re-sees it to close the loop. Layer 1 sensing → Layer 2 decision → Layer 3 action → Layer 1 verification.
Layer 1 · Data
Produces the ground truth of the shelf, tagged to fixture, zone, and timestamp. Per-scan-session uploads of high-resolution captures with bounding boxes.
Layer 2 · Intelligence
Turns ground truth into a compliance judgement, dimension by dimension, then decides what to do about it (triaged by SKU velocity and severity).
Layer 3 · Action
Turns the decision into the specific act that restores compliance, with the right person, the right SOP, and an evidence requirement.
Master data plane
The single source of truth for what compliance means on every fixture, in every store, for every role. Hourly sync; stale planograms block ticketing.
~t.compress(q:75)~t.toFormat(f:webp)/fynd-web/solutions/autri/autri-07-architecture-diagram.jpg)
Source diagram, 21-Apr-2026 brief. The dotted 'verified by next scan' arrow from Action back to Data is the contract that makes this self-verifying. Without it, this is just another ticket queue.
A ticket is closed only when the next AutRi scan confirms compliance on that fixture. Staff evidence alone is not closure. This is what makes the loop self-verifying and the compliance number trustworthy.
Step 1 · Scan
Robot runs scheduled sessions; captures shelf images per fixture and zone.
Output: tagged shelf images in the image repository.
Step 2 · Read
Reads the shelf against the active planogram version; detects presence, placement, facings, price, promotions.
Output: discrepancy list per fixture with evidence and confidence.
Step 3 · Decide
Triages each discrepancy by compliance dimension, severity, and SKU velocity; decides the action.
Output: triaged compliance alert with SOP mapping and route.
Step 4 · Act
Auto-creates a ticket with the shelf image, expected planogram view, and SOP; routes to the right role.
Output: ticket on the right person's PulsePoint app.
Step 5 · Verify
Staff evidence moves the ticket to pending verified. The next AutRi scan closes the ticket, or escalates.
Output: compliance confirmed, ticket closed, audit trail updated.
~t.compress(q:75)~t.toFormat(f:webp)/fynd-web/solutions/autri/autri-06-loop-diagram.jpg)
Scan → Read → Decide → Act → Verify, with a re-scan loop back to step 1.
~t.compress(q:75)~t.toFormat(f:webp)/fynd-web/solutions/autri/autri-08-end-to-end-flow.jpg)
End-to-end sequence — 17 numbered steps, from a scheduled scan to a verified closure. Every step has an owner and a handoff.
Each compliance dimension has a specific detection path, a specific severity decision, and a specific role it routes to. Velocity weighting and the active-promotion window override the default severity.
Value starts flowing before the full loop is live. Stages run in sequence, with Stages 2 and 3 overlapping through the middle of the build.
Stage 1 · Foundation
Align fixture, zone, SKU, and store IDs across AutRi, GMetri, and PulsePoint. Load grocery SKU catalogue and planogram library into GMetri. Ingest the promotion calendar.
Exit: Exit: data model signed off by all three teams; planogram live in GMetri; promotion calendar ingested.
Stage 2 · Intelligence
Tune GMetri across all six compliance dimensions. Write grocery triage rules in the Agentic Command Centre (dimension, severity, velocity, promotion-window) and the SOP mapping.
Exit: Exit: GMetri reliable at shelf level; triage rules approved by grocery ops leadership.
Stage 3 · Action in PulsePoint
Build the intake from Command Centre into PulsePoint. Configure grocery SOPs, evidence requirements, and the escalation matrix from Store Manager → Cluster Manager → State Head.
Exit: Exit: auto-ticketing live end-to-end; SOPs configured; escalation matrix live; audit trail captured.
Stage 4 · Pilot & scale
Run the end-to-end loop in two pilot grocery stores. Measure loop health weekly. Scale to the rest of the grocery estate after four consecutive weeks of steady KPIs.
Exit: Exit: pilot KPIs steady for four consecutive weeks — sign-off for chain-wide rollout.
Grocery is not fashion. The design choices below earn trust with store staff from day one and keep the inbox focused on what hits revenue.
An empty slot for a top-velocity SKU is not the same as a facing drift on a slow mover. The Command Centre weights triage by SKU velocity, not raw count of violations. Keeps the Store Manager's PulsePoint inbox focused on what hits revenue and customer experience.
Promotions drive a disproportionate share of grocery sales and customer trust. Price mismatches and missing end caps during an active promotion window are flagged Critical and go as push, regardless of the underlying compliance dimension. This lane has its own SLA.
Compliance signal is only meaningful after restocking has happened. AutRi scans are scheduled immediately after the primary restocking windows so alerts reflect post-restock reality. This single choice removes a large class of false alerts and earns trust with store staff from day one.
Nine targets, every one measurable. They are aspirational until the loop is live in pilot. Then they become the contract.
The loop only works if the staff trust it, the data stays fresh, and a robot outage doesn't stop the store.
Every claim above has a deeper artefact behind it.
Section-by-section build spec, asset pipeline, decisions.
AuditIndependent review, dimension scores, follow-ups.
SourceSaaket Chawali · 21-Apr-2026 · the original write-up.
Diagram3 layers + master data plane · the source artwork.
DiagramScan → Read → Decide → Act → Verify · with re-scan loop.
DiagramFrom scheduled scan to verified closure · per-role swimlane.
Fill out the form
Share your contact information to get started
Speak to an expert
A member of our sales team will get in touch with you