AI Agents · RETAIL-VISTA · Building

Retail Vista.

All-India spatial intelligence — site discovery, scoring, and a conversational workspace for the expansion team.

Building PLAN L4 Powered by  → Retail Vista
Spatial fabric
H3 grid
All-India · uniform hex resolution
Curated datasets
61
BigQuery · refreshed continuously
Scoring dimensions
10
wealth · demand · competition · ...
Specialist agents
4
Location · Catchment · Intelligence · Strategic
01 · The job to be done
The pain
Site selection runs on manual research, fragmented data, and individual judgement — weeks per location, inconsistent across teams, missed high-potential markets, and no visibility into cannibalisation risk.
What the agent does
Builds a single spatial fabric over India — H3 hex grid, 61 curated datasets, competitor map. Surfaces high-potential sites, scores them on 10 business dimensions, ranks by vertical-specific weights. The expansion team interrogates the map and the pipeline through a conversational workspace.
What humans still do
Expansion lead reviews the prioritised pipeline, signs off on shortlist visits, and takes the final lease decision. The agent does the discovery, scoring and analysis; the human owns commitment.
02 · How it works · the closed loop
Retail Vista · 4-stage opportunity engine + workspace Inputs (competitor map, demographics, JioMart density, property leads) feed a 4-stage pipeline that produces a ranked opportunity list, surfaced through a conversational workspace. Inputs Expansion lead vertical · target city capex envelope Competitor map stores · brands · density Demographics income · density · GDP JioMart density order hotspots · catchment Property leads listings · rates · footfall Substrate H3 hex grid · 61 datasets Pipeline · 4 stages + workspace 01 Candidate sourcing competitor ring expansion · whitespace density hotspots · property leads 02 Whitespace filtering isolate top-20% demand areas no-existing-coverage filter 03 10-dimension scoring wealth · density · GDP · footfall competition · accessibility · rates · ... 04 Vertical-weighted ranking Quick Commerce vs Fashion weights geographic diversity capping 05 Conversational workspace root orchestrator (Gemini 2.5 Pro) routes to 4 specialist agents: Location · Catchment · Intelligence · Strategic — each with dedicated tools 06 Human verification gates expansion team reviews shortlist · authorises field visits final lease commitment stays human KPIs moved Time-to-shortlist · selection accuracy · cannibalisation rate · expansion velocity Outputs Opportunity pipeline ranked, scored, vertical-weighted geographic diversity Site dossier 10-dim breakdown · competitor benchmark · catchment Live map answers natural-language queries over 61-dataset substrate Format Quick Commerce live · Fashion in build Autonomy L4 Agentic human reviews shortlist system runs the pipeline
Inputs (blue) → 4-stage pipeline + workspace (grey) → Outputs (green). Human verification gate (yellow) before lease commitment.
01
Candidate sourcing
Competitor ring expansion · whitespace detection · density hotspots · city seeding · property leads — surfaces a long-list per query.
02
Whitespace filtering
BigQuery spatial queries isolate the top-20% demand areas with no existing coverage.
03
10-dimension scoring
Wealth Index · Building Density · Demographics · District GDP · Spending Power · Footfall · Competitive Intensity · Accessibility · Property Rates · JioMart Order Density.
04
Vertical-weighted ranking
Composite score with vertical-specific weights. Quick Commerce emphasises demand + order density; Fashion emphasises footfall + complementary retail. Geographic diversity capping prevents cluster-heavy recommendations.
05
Conversational workspace
Root orchestrator (Gemini 2.5 Pro) routes natural-language queries to 4 specialist sub-agents — Location · Catchment · Intelligence · Strategic. Each runs its own tool stack against BigQuery, Google Maps APIs, and the H3 grid.
06
Human verification gates
Expansion team reviews the shortlist, signs off on field visits, takes the lease commitment. Agent does the discovery and analysis; human owns the decision.
03 · Underlying data
Data sourceClassificationKey entities
BigQuery (61 tables)MixedPopulation · Demographics · GDP · Wealth Index · Property Rates
Competitor mapMachine-CreatedStore locations · brand mix · density per hex
JioMart order dataMachine-CreatedOrder density · catchment · delivery time
Google Maps APIsMachine-ReadableRoutes · isochrones · POI · footfall proxies
Property leadsMixedListings · rates · landlord contacts

Refresh cadence: Substrate refreshed continuously; opportunity pipeline auto-runs on startup or on demand.

KPIs moved:

Time-to-shortlistSelection accuracy (year-1 sales vs plan)Cannibalisation rateExpansion velocity (sites opened per quarter)
04 · Design patterns used
Tool Calling
Each specialist sub-agent calls BigQuery, Google Maps, Routes API, and the H3 spatial library
Deep / Hierarchical
Root orchestrator routes queries to Location · Catchment · Intelligence · Strategic specialists
Wide Research
4-stage opportunity engine generates many candidates before scoring narrows to a shortlist
Context Graphs
H3 hex grid is the spatial graph; every entity (store, household, property) carries hex-level context
Observability
Every score breakdown is traceable back to the 10 dimensions and the underlying datasets
05 · Evals · how we know it works

Backtest opportunity scores against year-1 sales of recently opened Reliance Retail stores. Compare ranked recommendations vs the lease decisions actually taken; measure selection accuracy and missed-market cost.

Gates and thresholds.

  • Top-20% recommended sites match year-1 sales-per-sq-ft above the format median
  • Cannibalisation rate on recommended sites stays below 5%
  • Conversational workspace returns answers grounded in source datasets — every claim cites the dataset
06 · Linked platforms
Retail Vista
Platform home · live storefront for the agent
View platform →
Cortex
Consumes spatial / catchment context for store-level planning
View platform →
JCP
Delivery routing and catchment-aware fulfilment
View platform →