AI Agents · AI-CATALOGING · Live

AI Cataloging Agent.

Mannequin or product image → enriched catalog page with model-on photo, attribute set, and motion video. 5 hours, not 15 days.

Live SELL L4 Powered by → JCP Cataloging

Cycle compression

15 days → 5 hr

per shoot · same SKU set

Model library

AI-generated

ethnicity · age · size · gender · consistent face

Outputs per SKU

Photo + page + video

model-on shot · enriched attributes · motion video

Eval framework

External + internal

TIRA + Reliance reviewer audits

01 · The job to be done

The pain

Catalog production = 15 days per shoot. Doesn't scale across SKUs, seasons or brands. Output inconsistent — different studios, different models, different lighting. Attribute enrichment is manual and lossy. Motion video requires a separate production pass.

What the agent does

Takes a mannequin photo or product image. Generates a model-on shoot with consistent face from a brand-controlled model library (ethnicity, age, size, gender). Enriches the catalog page with structured attributes via VLM extraction. Produces a motion video of the same SKU. Runs in 5 hours per shoot at scale.

What humans still do

Brand sets the model library, attribute taxonomy, channel catalog spec, brand-tone rules. Catalog ops reviews flagged outputs (face drift, attribute confidence below threshold). Agent does the production; humans QA the exceptions.

02 · How it works · the closed loop

Inputs (blue) → 7-step pipeline (grey) with QA gate (yellow, auto-retry on fail) → channel-published outputs (green).

Garment understanding

VLM extracts category, cut, drape, colour, print and construction from the input mannequin / product photo.

Model selection

Picks a model from the brand-controlled library — ethnicity, age, size, pose constrained by brand spec.

Photoshoot generation

Generates the model-on shot with consistent face. Multi-angle, scene-aware lighting, brand-styled background.

Catalog enrichment

Structured attribute extraction mapped to channel-specific field schemas (TIRA, JioMart, AJIO each have their own taxonomy).

Video generation

4-6 second motion video — model walks / spins, with detail close-ups on construction features. PDP-ready.

QA gate · auto + human

Automated checks — face-consistency, attribute confidence threshold, channel-spec compliance. Auto-retry to step 03 on failure. Human QA reviews flagged exceptions only.

Publish to channels

Channel-native formatting and delivery: TIRA, JioMart, AJIO, Reliance Trends, Reliance Digital. CDN delivery for image + video assets.

03 · Underlying data

Data source	Classification	Key entities
Mannequin / product images	Mixed	Studio shots · vendor catalogue images
Brand model library	Human-Created	Approved model identities · ethnicity · age · size · pose
Attribute taxonomy	Machine-Readable	Colour · fabric · neckline · sleeve · pattern · construction
Channel catalog spec	Machine-Readable	TIRA / JioMart / AJIO / Reliance Trends field schemas

Refresh cadence: Per-SKU run · brand library and taxonomy refreshed seasonally.

KPIs moved:

Time-to-catalog (15 days → 5 hours)Attribute fill rateFace-consistency scorePDP conversion lift (with video vs without)

04 · Design patterns used

Tool Calling

VLM, image gen, video gen, channel-publish APIs, CDN — all tool-mediated

Multi-Agent Orchestration

Garment-understanding agent + Model-selection + Generation + Enrichment + QA agents collaborate

Wide Research

Multiple model + scene candidates explored before QA-gate narrows to publish set

Evals

External TIRA reviewer audit + internal eval framework + per-output QA checks · 3 layers

Observability

Per-SKU trace · model used · QA score · channel publish status · regen attempts logged

05 · Evals · how we know it works

Layer 1 — automated per-output QA: face-consistency score, attribute confidence, channel-spec validation; auto-retry on fail. Layer 2 — external TIRA Catalog Enrichment Evals: independent reviewer scores accuracy, completeness and brand alignment on a sample. Layer 3 — internal Reliance reviewer rubric for ongoing regression detection.

Gates and thresholds.

Face-consistency score above brand threshold per SKU
Attribute fill rate ≥ 90% per channel-spec mandatory fields
External reviewer accuracy ≥ baseline manual catalog quality