CartMind — Personalized Shopping Decision Engine

Inspiration

Online shopping gives users thousands of products, but finding the right product at the right effective price is still difficult.

Traditional comparison tools mainly compare listed prices. They often do not understand what the user actually needs, whether a specific product variant matches those requirements, or which available offers the user is eligible for.

We wanted to build something different: instead of asking users to compare dozens of products themselves, CartMind helps them make the decision.

What We Built

CartMind is a personalized shopping decision engine that understands natural-language shopping requests and converts them into structured shopping intent.

For example:

"I need a laptop under ₹50,000 with 16GB RAM and SSD for coding."

CartMind identifies the important requirements, finds matching products, checks hard constraints, evaluates eligible offers, calculates the effective price, and ranks the best options.

The goal is not simply to find the cheapest product, but to answer:

What should I buy, why does it fit my needs, and what will I actually pay?

How We Built It

CartMind follows a hybrid AI + deterministic architecture.

AI — Understand

We use Amazon Bedrock to understand natural-language shopping queries and extract structured ShoppingIntent.

The AI handles tasks where language understanding is important, such as identifying:

• Product category
• Budget
• Required specifications
• User preferences
• Shopping intent

Java — Decide

Deterministic Java services handle decisions that should be predictable and explainable:

• Product and variant matching
• Hard constraint filtering
• Offer eligibility
• Effective price calculation
• Product ranking
• Data confidence

This prevents the LLM from arbitrarily deciding numerical product rankings or prices.

AI — Explain

After the deterministic decision process, AI can help explain the recommendation in a user-friendly way.

The overall principle is:

AI = Understand → Java = Decide → AI = Explain

Effective Price

One of CartMind's key ideas is that the displayed product price is not always the price that matters to a specific user.

CartMind introduces an effective price layer that considers eligible offers and user-provided wallet information such as bank, card type, and membership.

For example:

Listed Price − Eligible Discount = Effective Price

This allows two products with similar listed prices to be compared based on what the user may actually pay.

Product & Variant Matching

CartMind separates products from their variants so that specifications are not accidentally mixed.

For example, an 8GB/256GB variant should not be treated as equivalent to a 16GB/512GB variant.

This improves the accuracy of recommendations and prevents misleading comparisons.

Challenges We Faced

One of our biggest challenges was deciding what should be handled by AI and what should remain deterministic.

It was tempting to let an LLM perform the entire recommendation process. However, pricing, eligibility, filtering, and ranking require consistency and explainability.

We therefore designed a hybrid architecture where Amazon Bedrock focuses on language understanding while deterministic backend logic handles business-critical decisions.

Another challenge was designing the system so that stale or uncertain product and offer information does not appear as completely reliable. We therefore consider data freshness and confidence as part of the decision process.

What We Learned

Building CartMind taught us that adding AI to a product is not enough.

The important question is:

Where does AI actually provide value, and where should deterministic engineering take control?

We learned to combine LLM capabilities with traditional software engineering instead of making the LLM responsible for every decision.

We also learned the importance of designing around the user's actual decision rather than simply presenting more information.

What's Next

Our initial focus is on validating the core decision engine before expanding to many categories and shopping platforms.

Future improvements include:

• More product sources
• More categories
• Better offer intelligence
• Improved product matching
• Stronger freshness and confidence signals
• Personalized recommendations based on user preferences

Our long-term vision is to make CartMind a personalized shopping decision engine, rather than another generic price-comparison tool.

Built With

• ai
• amazon
• amazon-web-services
• bedrock
• e-commerce
• generative
• recommendation