Retail New Item Evaluation

In this demo, a supplier submits a new grocery product and AI on Cloudera evaluates it against the retailer’s existing assortment in about a minute. CLIP image-and-text embeddings and OpenSearch vector search find the closest look-alike products on the shelf. A rules-based engine then sets an authorize, modify, or decline verdict, and three CrewAI agents explain it with a cannibalization risk assessment, replacement candidates among underperforming SKUs, and best-, expected-, and worst-case revenue projections. Progress streams live as each step completes, and reviewers can ask follow-up questions about any recommendation. 
Watch the demo

1. Multimodal visual similarity search. CLIP embeds each product's image and text together, and OpenSearch k-NN search finds look-alike products already on the shelf, so the system compares what products look like, not just their descriptions.

2. Auditable, deterministic verdicts. The AUTHORIZE, MODIFY, or DECLINE decision comes from a transparent rule based on overlap and category saturation, while AI agents write the cannibalization, market, and financial analysis that supports it. The same product always gets the same verdict.

3. End-to-end on Cloudera. It deploys as a one-click AMP in a Cloudera AI Workbench. Agents call open-weight models on Cloudera AI Inference, and retail data is queried as Iceberg tables through Cloudera Data Warehouse Impala.

• Category management: screen new items before line reviews. Each supplier proposal is compared visually and semantically to the current assortment, flagging items likely to cannibalize existing SKUs and highlighting white-space opportunities where the category has room.

• Assortment planning: make room for winners. When a new item overlaps the shelf, the system identifies declining, clearance, seasonal, or bottom-shelf SKUs it could replace and calculates the net revenue impact of the swap.

• Merchandising operations: triage the queue at scale. Batch evaluation processes a backlog of proposals at once, so buyers can spend their time on the submissions that need human judgment.

• Category finance: see the numbers before committing. Best-, expected-, and worst-case revenue projections, along with a suggested retail price, placement, and rollout plan, help teams weigh each decision against category margins.

• Supplier and vendor management: give clear, consistent feedback. Modify and decline verdicts come with specific feedback suppliers can act on before resubmitting, and vendor performance data such as fill rate and on-time delivery informs every evaluation.

• Governance and audit: explain every decision. Verdicts follow a transparent rule, so the same product in the same category gets the same outcome, and follow-up Q&A lets reviewers question the reasoning behind any recommendation.

• Beyond grocery: reuse the pattern. Multimodal similarity search combined with rules-based decisions and agent-written analysis carries over to private-label development, apparel and beauty line planning, and any category where how a product looks on the shelf matters as much as its specs.

Reference architecture on Cloudera AI. The Application pod in the AI Workbench project runs the API and UI, CLIP and an embedded OpenSearch k-NN index. Sales, category and vendor data come from Iceberg tables through an Impala Virtual Warehouse in Cloudera Data Warehouse, and the three agents run on an open-weight model served by Cloudera AI Inference. Numbers 1 to 7 follow one evaluation: submit, embed, find look-alikes, query Impala, compute the verdict, run the agents, validate and save.