Architecture
TalkWise AI Logo

TalkWise AI

Smart support, right when it’s needed most.

Inspiration

Live customer support is chaotic. Agents are under pressure to resolve issues fast—without always having the right information in front of them. We wanted to build a tool that brings AI-powered assistance into live calls, helping agents respond quickly, accurately, and calmly.

TalkWise AI is that tool: a real-time, GenAI-powered voice assistant designed to make customer interactions smoother, smarter, and more efficient.

What it does

TalkWise AI listens during live customer conversations (via chat or speech-to-text) and delivers real-time GenAI suggestions—everything from relevant responses and clarifications to helpful nudges and summaries. It also stores interactions for later analysis.

Core Features:

Real-time message handling over WebSocket
Intelligent responses generated via Amazon Bedrock (Claude v2)
Durable chat history with DynamoDB
Clean, fast, and reactive React.js UI for live agents
Fully serverless backend on AWS Lambda

How we built it

Stack Breakdown:

Layer	Tech Used
Frontend UI	React.js + Vercel
Real-time Comm	Amazon API Gateway (WebSocket)
Backend Logic	AWS Lambda
AI Intelligence	Amazon Bedrock (Anthropic Claude v2)
Data Store	Amazon DynamoDB
Infra-as-Code	AWS CDK (TypeScript)

Everything was wired up with WebSocket events to allow streaming of real-time messages, and Lambda functions process each message before passing it to the Bedrock foundation model. The final response is routed back to the client over WebSocket.

Challenges we ran into

WebSocket Event Parsing: One wrong character and nothing connects.
IAM and Bedrock access: Configuring model permissions took... several attempts.
CDK Deployment Errors: Mismatched paths, handler names, and missing .js files gave us heartburn.
Real-time UX: We wanted responses to feel “instant”—even when they weren’t.

Accomplishments that we're proud of

Fully functional, serverless, real-time GenAI app in under a week.
Live-deployed and publicly accessible via talkwise-ai.vercel.app
Seamlessly integrated Amazon Bedrock Claude v2 for production-quality responses.
Built a calm and trustworthy GenAI experience—not a chatbot that yells.

What we learned

Bedrock’s Claude v2 works well with short-form, real-time prompts.
WebSockets require very specific handling of state and connections.
CDK is powerful—but you must respect its folder structure and asset handling.
A good AI experience is about timing, tone, and simplicity.

What's next for TalkWise AI

Amazon Transcribe Integration – Convert real-time audio to text.
Sentiment Detection – Adapt suggestions based on customer tone.
Post-call Summaries – Automatically generate wrap-ups for agents.
Multilingual Support – For global agent teams.
Responsible AI – Logging, auditing, and bias mitigation.

Responsible AI

TalkWise AI was built with Responsible AI principles in mind:

Transparency: All AI-generated responses are labeled and stored.
Accountability: Logged message history ensures auditability.
Fairness: Prompts are stripped of bias indicators, and model selection is intentional.
Security: All communication is encrypted via AWS services. No user PII is stored.