Skip to content
Flagship Case Study

Product Intelligence Engine

AI-powered product discovery built on enterprise data.

Full-stack RAGProduction-readyIndependently built

An intelligent product recommendation system built in React and powered by AI. Designed to help users find the right products for customers in seconds using enterprise product data. Demonstrated live to executive leadership who backed the project for production.

Recommendation workflow, from prompt to results
< 7s
Live executive demonstration
CEO + CFO
Endorsed for production after the demo
Backed
Products the system was designed to search
42,000+
The Problem

Manual product lookup doesn't scale

Finding the right products for customers often requires manually searching large product catalogs or relying on AI tools that lack access to company-specific product data. It's a slow and repetetive process that doesn't scale well and is difficult to automate. The effort adds up on every request, taking away time from actually helping customers.

The Solution

An instant, intelligent recommendation engine

I designed and built a retrieval-augmented recommendation engine in React that returns relevant product recommendations in less than seven seconds. It searches internal product catalogs semantically and transforms a manual lookup process into one that is instant and reliable.

Semantic search

Finds relevant products based on usage and context rather than exact keyword matches.

Context-aware generation

Builds recommendations using retrieved product data to improve accuracy and reliability.

Production-ready performance

Delivers product recommendations in a matter of seconds and is suitable for live customer interactions.

Architecture

How a request flows through the system

The system retrieves products from the catalog that are similar to the user's request, then uses AI to rank, explain, and recommend the top matches.

Sales repUserEnters a customer's requirements or use-case in plain language.
Client UIReact FrontendCaptures the request and renders ranked product recommendation cards.
OrchestrationPython APICoordinates the embedding & retrieval workflow.
Query embeddingEmbedding GenerationEncodes the user's request into a vector to prepare for semantic search.
Vector retrievalPineconeReturns the most semantically similar product candidates from the catalog.
Context-based generationGeminiReasons over the retrieved products to identify the closest matches.
ResponseSmart RecommendationsDelivers the top results with explanations and confidence scores.
Engineering

Decisions & technical challenges

Getting the system to return products wasn't particularly difficult, and the first versions already worked fairly well. Most of the engineering effort went into improving recommendation quality, ensuring consistent results, and keeping the response times low enough for real-world use.

Evidence-based generation

Recommendations had to come from an actual catalog, not the model's imagination. Forcing the model to only use products retrieved from the catalog with strict constraints kept the output accurate and trustworthy, minimizing hallucinations.

Semantic retrieval quality

Customers rarely describe products using the exact same language found in a catalog or database. Embeddings allows the system match on overall meaning, helping surface relevant products even when terminology differs.

Minimizing latency

Balancing retrieval quality against generation time proved to be a major challenge. Improving response times required optimizing every stage of the pipeline, from embedding generation to retrieval, prompt design, and caching techniques.

Prompt design

Carefully designed system prompts force the model to compare products using the same evaluation criteria each time, keeping outputs consistent with reduced variability between requests.

Relevance ranking

Returning relevant products isn't enough. The most useful ones have to come in order, and include confidence scores to increase trust.

Built for production

Users won't always search with perfect terminology, so the system was built to handle ambiguous requests while remaining simple enough for non-technical people to use with minimal training.

Stack

Core Technologies

01

User Interface

  • React
  • Responsive UI
  • Recommendation Cards
  • Guided Input Forms
  • Freeform Text Prompts
02

Orchestration Layer

  • Python
  • FastAPI
  • REST API
  • Request Orchestration
03

AI & Retrieval

  • Gemini API
  • Pinecone
  • Semantic Search
  • RAG
  • Embedding Generation
04

Engineering Practices

  • Prompt Engineering & Design
  • Grounded Generation
  • Relevance Ranking
  • Latency Optimization
Impact

Results & impact

Recommendation workflow, from prompt to results
< 7s
Live executive demonstration
CEO + CFO
Endorsed for production after the demo
Backed
Products the system was designed to search
42,000+
  • Replaces a manual lookup process with instant AI-assisted recommendations.
  • Returns ranked product recommendations with explanations and confidence scores.
  • Supports both freeform prompts and guided input forms for different types of users.
  • Demonstrated live to executive leadership and approved for production.
  • Built to provide consistent recommendations across a large evolving product catalog.
  • Full-stack ownership: designed and built independently across React, Python, Pinecone, and Gemini.
Live Demo

Try it by request

See how the recommendation engine works with realistic sample data.

Live demo · In development

Interactive Demo

I'm currently building a sandbox version of the Product Intelligence Engine using synthetic data. Over time, I'll be adding additional datasets for different industries to show how the same retrieval and recommendation architecture can be applied to different industries.

Request demo accessPassword-protected sandbox in development

Independently designed and built by Caleb Kennedy. Business-specific details and datasets are kept entirely confidential.

Let's connect

Let's build something intelligent.

I'm open to AI/ML and data engineering roles and collaborations. If you have a problem that needs solving, I''d love to hear about it.