Kalakriti: AI-Powered Marketplace for Indian Artisans 🎨

Empowering 7M+ Indian artisans with Gemini 3 AI—voice-to-listing in 12 languages, AI photo editing, and agentic shopping. Bridging the digital divide, preserving cultural heritage.

🌟 Inspiration

India is home to 7 million artisans creating exquisite handicrafts—Madhubani paintings, Warli art, Kalamkari textiles, and intricate pottery passed down through generations. Yet, most of these artisans:

📵 Cannot use smartphones effectively due to digital illiteracy
🗣️ Speak only regional languages (Hindi, Bengali, Tamil, etc.)
📸 Lack professional photography skills to showcase their art
💰 Lose 40-60% margins to middlemen who exploit their isolation

We asked ourselves: What if AI could become the bridge between traditional artisans and the global digital marketplace?

Kalakriti was born from this vision—a platform where an artisan in rural Bihar can speak in Hindi, describe their Madhubani painting, and have AI create a professional e-commerce listing complete with enhanced photos, cultural context, and smart pricing.

🎯 What We Built

Kalakriti is a full-stack AI-powered marketplace that leverages Gemini 3 to solve the unique challenges faced by Indian artisans.

Core Features Powered by Gemini 3

Feature	Gemini 3 Model	Description
AI Auto-Cataloging	`gemini-3-flash-preview`	Multimodal product listing from voice + images
Professional Photo Studio	`gemini-3-pro-image-preview`	AI background removal & image enhancement
Agentic Personal Shopper	`gemini-3-flash-preview`	Intelligent tool selection & execution
Product-Specific Chatbot	`gemini-3-flash-preview`	Context-aware Q&A for each product
Cultural Context Engine	`gemini-3-flash-preview`	Generates heritage information for products
Smart Gift Recommendations	`gemini-3-flash-preview`	AI-curated gifts based on recipient & occasion

Additional AI Integrations

Feature	Technology	Purpose
Multilingual Support	Bhashini API (`ai4bharat/indictrans-v2`)	Translation across 12 Indian languages
Voice Input	Bhashini API (`ai4bharat/conformer`)	Speech-to-text for Hindi & regional languages
Smart Logistics	ShipRocket REST API	Automated shipping & delivery tracking

🏗️ How We Built It

Architecture Overview

┌─────────────────────────────────────────────────────────────────────┐
│                         KALAKRITI ARCHITECTURE                       │
├─────────────────────────────────────────────────────────────────────┤
│                                                                      │
│  ┌──────────────┐     ┌──────────────┐     ┌──────────────┐        │
│  │   Next.js    │     │   Gemini 3   │     │   Bhashini   │        │
│  │   Frontend   │────▶│   Flash &    │────▶│   API        │        │
│  │   + API      │     │   Pro Image  │     │   (NLP)      │        │
│  └──────────────┘     └──────────────┘     └──────────────┘        │
│         │                    │                    │                 │
│         ▼                    ▼                    ▼                 │
│  ┌──────────────┐     ┌──────────────┐     ┌──────────────┐        │
│  │   Prisma     │     │   Custom     │     │  ShipRocket  │        │
│  │   ORM        │────▶│   AI Agent   │────▶│  Logistics   │        │
│  │              │     │   Framework  │     │              │        │
│  └──────────────┘     └──────────────┘     └──────────────┘        │
│         │                                                           │
│         ▼                                                           │
│  ┌──────────────────────────────────────────────────────────┐      │
│  │              Google Cloud SQL (PostgreSQL)                │      │
│  └──────────────────────────────────────────────────────────┘      │
│                                                                      │
│  Deployment: Google Cloud Run (Serverless)                          │
└─────────────────────────────────────────────────────────────────────┘

Tech Stack

Layer	Technology
Frontend	Next.js 15, React 19, TailwindCSS 4
Backend	Next.js API Routes, TypeScript
Database	PostgreSQL on Google Cloud SQL
ORM	Prisma
AI/ML	Gemini 3 Flash, Gemini 3 Pro Image, Bhashini
Deployment	Google Cloud Run (Serverless)
Logistics	ShipRocket API
Maps	Leaflet + OSRM

Custom Agentic AI Framework

We built a custom agent framework (no LangChain dependency) that enables intelligent tool selection:

// AI Agent Architecture
class AIAgent {
  private tools: AgentTool[]  // 5 specialized tools

  async processQuery(userMessage: string, context: AgentContext) {
    // Step 1: Gemini 3 analyzes intent & selects tools
    const toolSelection = await this.selectTools(userMessage, context)

    // Step 2: Execute selected tools with extracted parameters
    for (const { toolName, parameters } of toolSelection.tools) {
      const result = await tool.execute(parameters, context)
      toolResults[toolName] = result
    }

    // Step 3: Generate natural language response
    return await this.generateResponse(userMessage, toolResults)
  }
}

Available Agent Tools:

search_products - Advanced product search with filters
recommend_gifts - AI-curated gift recommendations
analyze_price_range - Market price analysis
get_cultural_context - Heritage information retrieval
get_personalized_recommendations - User history-based suggestions

🧠 Gemini 3 Integration Deep Dive

1. Multimodal Auto-Cataloging

Artisans can create listings by simply:

🎤 Speaking in their native language
📸 Uploading a photo of their craft

Gemini 3 Flash processes both inputs to generate:

Professional title & description
Relevant tags for searchability
Suggested pricing based on market analysis
Cultural significance context

// Multimodal product creation
const prompt = `
Analyze this handcrafted product image and voice description.
Generate a professional e-commerce listing with:
- Title, Description, Tags
- Material & Technique detection
- Cultural significance
- Suggested price range
`
const response = await model.generateContent([prompt, imageData, audioText])

2. AI Photo Studio (Gemini 3 Pro Image)

Rural artisans often photograph products in poor lighting with cluttered backgrounds. Our AI Photo Studio:

Enhances image quality while preserving authenticity
Removes/replaces backgrounds with professional settings
Maintains 100% product authenticity (guardrails enforced)

// Responsible AI guardrails
const prompt = `
STRICT REQUIREMENTS:
1. KEEP THE PRODUCT EXACTLY AS IS - preserve all original details
2. DO NOT change product colors, patterns, or textures
3. ONLY enhance: lighting, background, clarity
4. Maintain artisan authenticity at all costs
`

3. Agentic Personal Shopper

Our agent understands complex queries like:

"Find me a gift for my colleague under ₹2000, something traditional from Rajasthan"

Gemini 3 Flash:

Parses intent → Gift recommendation
Extracts parameters → recipient: colleague, budget: 2000, origin: Rajasthan
Selects tool → recommend_gifts
Generates response with curated products

🚧 Challenges We Faced

1. Balancing AI Enhancement vs. Authenticity

Challenge: AI image models tend to "improve" products by changing colors or adding details—destroying the handmade authenticity that makes each piece unique.

Solution: We implemented strict prompt guardrails that explicitly prohibit modifications to product characteristics:

// Guardrail example
"KEEP THE PRODUCT EXACTLY AS IS - do not modify the product itself in any way"
"Preserve all original details, colors, patterns, and textures"

2. Multilingual Complexity

Challenge: India has 22 official languages with different scripts (Devanagari, Tamil, Bengali, etc.). Direct translation often loses cultural nuance.

Solution: We integrated Bhashini API (Government of India's NLP platform) which is specifically trained on Indian languages and cultural context, then used Gemini 3 for semantic understanding.

3. Building Agentic AI Without Wrappers

Challenge: LangChain and similar frameworks add complexity and dependencies. We needed a lightweight, customizable solution.

Solution: Built a custom agent framework with:

Tool definitions with JSON schemas
Gemini 3-powered tool selection
Fallback intent detection for reliability
Structured response generation

4. Image Processing at Scale

Challenge: Base64 encoding large images for Gemini 3 Pro Image API calls created memory and latency issues.

Solution:

Implemented URL-based image fetching
Added MIME type validation
Chunked processing for large batches

📚 What We Learned

Technical Insights

Gemini 3's multimodal capabilities are exceptional for real-world applications where users provide mixed inputs (voice + image + text)
Custom agent frameworks can outperform wrapper libraries when you need precise control over tool selection and execution
Prompt engineering for image generation requires explicit "do not" instructions to prevent unwanted modifications
Indian language AI requires specialized models—Bhashini + Gemini 3 combination works better than Gemini alone for regional languages

Product Insights

Digital illiteracy is a design problem, not a user problem. Voice-first interfaces can unlock massive underserved markets
Cultural context matters for e-commerce. A Madhubani painting is not just "wall art"—it's heritage
AI can preserve authenticity when properly guardrailed, becoming an enabler rather than a homogenizer

🔮 Future Roadmap

Feature	Technology	Status
AR Product Preview	Gemini 3 + WebXR	Planned
Artisan Story Videos	Gemini 3 Video	Planned
Real-time Translation Chat	Gemini 3 + Bhashini	In Progress
ONDC Integration	Open Network Protocol	Planned
Blockchain Authenticity	Hyperledger	Research

🏆 Impact Metrics

Metric	Value
Target Artisans	7+ Million in India
Languages Supported	12 (Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Odia, Punjabi, Urdu, English)
Margin Improvement	40-60% (eliminating middlemen)
Listing Creation Time	Reduced from hours to minutes
Photo Quality	Professional studio-grade from smartphone photos