Sitaara

Intelligent Customer Support Automation

Comment

🧠 Sitaara – Hackathon Project Story

🚀 Inspiration

After completing my AWS Solutions Architect Certification, I was eager to apply serverless best practices in a real-world setting. I’ve always been fascinated by how customer support could be automated using AI, and that’s what sparked Sitaara — an AI-powered platform that automates issue resolution using intent recognition and tool execution.

My goal was to build something truly modular and scalable — and Sitaara reflects that vision.

💡 What it does

Sitaara is an AI-first customer support platform that:

  • Understands customer issues through natural language
  • Identifies the correct intent using vector search (OpenSearch) and LLM
  • Follows a predefined workflow of steps to resolve issues
  • Dynamically executes tools (actions) to automate actions like data lookups, notifications, or validations
  • Provides an AI-driven chat interface built with AWS Cloudscape
  • Stores all interactions for later auditing, insights, and learning

🛠️ How we built it

We used a completely serverless architecture on AWS:

  • Backend (All Serverless):

    • Lambda Functions handle all logic: CRUD, AI processing, intent recognition, tool execution
    • API Gateway for HTTP endpoints
    • SQS Queues for decoupled message orchestration
    • OpenSearch Serverless for intent vector matching
    • Bedrock (Anthropic Claude) for AI conversations
    • DynamoDB and MongoDB Atlas for persistence

  • Frontend:

    • Built in React using AWS Cloudscape Design System
    • Hosted on S3, served globally via CloudFront
    • Uses presigned S3 URLs for secure media uploads
    • Live demo site

Architecture reference: See Architecture

🧱 Challenges we ran into

  1. CORS Issues: Spent hours debugging why frontend couldn’t talk to API Gateway. Later realized I didn’t enable Lambda Proxy Integration, which caused events to be misformatted and headers to fail.

  2. Voice + WebSocket Challenge:
    I built a voice assistant using Twilio, Transcribe, and Polly for real-time AI support. Initially tried using API Gateway WebSockets, thinking Lambda maintained persistent connections, but learned it only handles discrete events. This led me to shift from chat/live support to structured support workflows — better suited for automating repetitive, high-volume support cases.

  3. Presigned S3 Uploads: Generating correct policies and matching content types with upload flow took tuning.

  4. Intent Routing: Mapping user queries to intents using vector embeddings was tricky — needed to tune confidence thresholds and fallback logic.

🎉 Accomplishments that we're proud of

  • Designed a production-grade serverless architecture with real-time AI integration
  • Seamlessly integrated Cloudscape UI to match AWS native UX
  • Created a working dynamic tool execution engine with custom scripts and parameter passing
  • Ran realistic healthcare test cases with ambiguity handling, and escalation paths

📚 What we learned

  • Deepened understanding of AWS Lambda, especially integration with:

    • SQS
    • API Gateway
    • OpenSearch Serverless
    • Bedrock and presigned S3 uploads

  • Realized the value of event-driven design and the flexibility it brings in building complex workflows

  • Learned the hard way about the importance of CORS and event formatting when integrating frontend and Lambda through API Gateway

  • Learned the importance of observability in event-driven serverless architectures using CloudWatch logs, metrics, and structured tracing to debug and monitor distributed flows.

🔮 What's next for Sitaara

  • Add authentication via Cognito or Firebase
  • Enable human escalation workflows for unresolved intents
  • Improve case analytics dashboard for admin review
  • Add multi-language support and broader industry use-cases
  • Explore cost-based optimization with Lambda Power Tuning and Cold Start improvements
Share this project:

Updates