New Product Development Assistant

Launch your new 'hero' product NOW. Validated by market trends and your sales data!

Comment

New Product Development Assister

Slides Video Link

Inspiration

This project was inspired by the challenge small Shopify merchants face in keeping customers engaged over time. Retention is crucial — acquiring new customers costs up to five times more than retaining existing ones, yet many stores lose buyers due to stale or misaligned product offerings. NPD Assistant empowers merchants to make data-driven decisions on what to launch next by analyzing their store’s sales patterns, customer preferences, and market trends. It works through three AI agents — Market Research, Sales Analysis, and Strategic Advisor — collaborating in real time to recommend the most relevant and profitable next product or color.

What it does

Our system deploys three specialized AI agents that collaborate to solve product development decisions:

Market Research Agent searches the web for current trends, popular colors, and consumer preferences specific to your target demographic.

Sales Analysis Agent analyzes your internal data using 6 custom tools to identify top performers, underperforming products, color trends, and customer patterns.

Strategic Advisor Agent synthesizes both insights and generates prioritized recommendations with clear rationale and expected business impact.

How it works:

  1. Upload sales data + choose NPD strategy (retain audience, improve products, new category, or new audience)
  2. Select target demographic (men, women, kids)
  3. AI agents collaborate using real tools (web search, data analysis)
  4. Receive prioritized recommendations with quick wins and long-term strategies

Example output: "Launch olive green tech pants—trending in market research + your current tech pants have 15% repeat purchase rate = high-probability success."

How we built it

Tech Stack:

  • AWS Bedrock with Claude Sonnet 4 for AI agents
  • Python/FastAPI backend with custom agent orchestration
  • React + Vite dashboard for real-time agent activity
  • Pandas for sales data analysis
  • Tavily API for web search

Architecture: Three specialized agent classes, each with conversation history and tool access. The orchestrator coordinates sequential execution: Market Research → Sales Analysis → Strategic Advisor synthesis.

Tool Calling Implementation: Used Bedrock's Converse API with native tool support. Agents iterate (up to 10-15 loops) using tools like web_search, get_top_products, get_color_performance, etc. Each tool call is logged for transparency.

Agent Loop Pattern:

while iteration < max_iterations:
    response = bedrock.converse(messages, tools)
    if tool_use: execute tools, continue
    if end_turn: return insights

Frontend: React Query polls for updates every 5 seconds, displaying agent activity logs in real-time. Beautiful gradient UI (purple/indigo) with charts and expandable recommendations.

Key Innovation: True multi-agent collaboration where each agent specializes and contributes unique insights—not just one AI with multiple prompts.

Challenges we ran into

Agent conversation state: Managing multi-turn tool calling was complex. Solution: Carefully structured conversation history with proper toolResult formatting.

Iteration control: Agents kept exploring indefinitely. Solution: Max iteration limits (10-15) with clear stopping conditions.

Structured output: Getting consistent JSON recommendations. Solution: Explicit schema in system prompts + resilient parsing.

Real-time UX: Backend runs synchronously but users want live updates. Solution: React Query polling + activity logs create real-time feel.

Tool output size: Some queries returned massive data exceeding token limits. Solution: Smart summarization for logs while passing full data to agents.

CSV data quality: Real data is messy. Solution: Robust validation and cleaning in tool initialization.

Demo reliability: External APIs can fail. Solution: Mock data fallbacks + health checks + comprehensive testing.

Accomplishments that we're proud of

True multi-agent collaboration - Three specialized agents with distinct expertise working together like a real business team

Production-quality architecture - Modular, scalable, well-documented system ready for real deployment

Beautiful dashboard - Modern UI showing agent reasoning in real-time with charts and visualizations

Advanced tool calling - Mastered Bedrock's tool API with 7+ custom tools and proper conversation management

Actionable insights - Recommendations that actually work, backed by cross-referenced market and sales data

Comprehensive documentation - QUICKSTART, demo scripts, and implementation notes

Fast performance - Complete analysis in 2-5 minutes despite multi-agent coordination

Resilient design - Graceful degradation, error handling, and mock data fallbacks ensure reliability

What we learned

Multi-agent systems are powerful: Specialized agents collaborating produce better results than single models. Each agent focuses on its expertise, then synthesis creates insights no individual agent could achieve.

Tool calling is transformative: Giving AI access to real tools (search, data analysis) makes it exponentially more capable than prompts alone.

Prompt engineering is both art and science: We iterated dozens of times. Context, examples, and clear structure matter. System prompts define identity, user prompts provide dynamic context.

Real-time UX requires creativity: Polling with React Query works great for perceived real-time updates. Activity logs keep users engaged during AI thinking time.

Data quality is critical: Clean your data early. Pandas errors mid-analysis are frustrating—validate and sanitize upfront.

Documentation is development: Writing docs early clarified our architecture and made the project more maintainable.

Resilience beats optimization: For demos and MVPs, working reliably is more important than peak performance. Mock data fallbacks and error handling prevent embarrassing failures.

We can build anything: We tackled an ambitious project with new technologies (Bedrock, tool calling, multi-agent orchestration) and shipped something we're genuinely proud of.

Built with ❤️ using AWS Bedrock, Claude Sonnet 4, FastAPI, and React
Hackathon Project • October 2025

Built With

Share this project:

Updates