Build AWS AI Demo Agent

AI Demo Agent: Project Brief
7 Nov 2026 · @Devon
Goal and scope
A "Get demo" button starts a live video call where an AI agent runs a tailored demo by voice, answers questions from your FAQ, and passes qualified leads to your team.
• Built from your 30+ recorded demos (30-40 min each) and your FAQ
• Runs on AWS, built with Claude Code, embedded in your software
• Available 24/7; similar in idea to Hyper AI (gethyper.ai)
• Not in version 1: agent operating a live copy of your software, phone calls, talking avatar
Customer journey
1. Customer clicks "Get demo" and the call connects in seconds.
2. Agent asks role, organization and the problem to solve.
3. Agent picks 3-5 chapters from your 30+ demos, turning 30-40 minutes into a tailored 10-15 minute demo.
4. Agent plays each chapter and narrates; the customer can interrupt ("go back", "how does billing work?").
5. Agent answers from your FAQ and demo transcripts; if the answer isn't there, it flags a person.
6. Agent books a sales call, sends a recap email, and logs the summary to your CRM.
Who does what
Claude writes the code, a developer deploys it on your AWS account, and your team supplies content and decisions.
Claude (in Claude Code)
• Project scaffold, voice agent and agent tools (pick chapters, play, pause, search FAQ, book meeting, hand off)
• Pipeline that transcribes the demos, splits them into 2-4 minute chapters and tags them
• FAQ retrieval, "Get demo" widget and call screen
• CRM, calendar and email integration code
• AWS infrastructure as code (AWS CDK), tests and documentation
Developer
• Prepare the AWS account and environments; request Claude model access on Amazon Bedrock
• Deploy, and review Claude's code before it goes live
• Store credentials in AWS Secrets Manager; embed the widget in your software
• Monitoring, cost limits, backups and security review
• Privacy and healthcare compliance setup (AWS BAA, encryption, access logs) if patient data is involved
Your team
• Provide the 30+ demo recordings and the FAQ
• Review chapter tags; decide the agent's name, voice, tone and what it must never say
• Choose the CRM and calendar and give access
• Run test calls and report what feels wrong
• Approve legal text (recording consent, privacy notice) and set handoff rules
AWS services for the live video call
Amazon Chime SDK is the core; speech, Claude, video delivery and call servers surround it.
Service
Role in the call
Amazon Chime SDK (Meetings)
Live audio and video room for customer and agent (web, iOS, Android); data messages tell the browser "play chapter 3"
Chime SDK Media Pipelines (or Kinesis Video Streams)
Sends the customer's audio to the agent; records calls
Amazon Transcribe (streaming)
Speech to text
Amazon Bedrock (Claude)
The agent's brain: chooses chapters, calls tools, answers
Amazon Polly
Text to speech, streamed into the call
Amazon S3 + CloudFront
Store and deliver demo videos; the browser plays chapters locally, in sync with the agent
AWS Elemental MediaConvert
Converts recordings to streaming formats
Amazon ECS on AWS Fargate
One agent worker per live call (too long-running for Lambda)
Amazon API Gateway + AWS Lambda
Starts the call, issues join tokens, receives webhooks
Amazon DynamoDB
Live session state and call logs
Amazon Route 53 + Certificate Manager
Domain and HTTPS
Optional: Cognito, Nova Sonic, avatar
Cognito for customer sign-in; Nova Sonic (speech-to-speech, lower delay) to test later; a talking avatar is third-party, not AWS
AWS services for knowledge, integration and security
Area
Services
FAQ and answers
Bedrock Knowledge Bases + OpenSearch Serverless (retrieve from your FAQ and transcripts); Bedrock Guardrails (blocks unsupported or off-topic replies)
Demo library pipeline
S3 upload, EventBridge + SQS, then Step Functions + Lambda: Transcribe (batch), chapter, tag with Claude, store
Follow-up and CRM
Amazon SES (recap email); API Gateway + Lambda (CRM and calendar); Secrets Manager (credentials)
Security
KMS (encryption), IAM (least access), WAF (bot protection), CloudTrail (audit log)
Operations
CloudWatch + X-Ray (delay, errors, cost per call); AWS CDK + CodePipeline or GitHub Actions (deployment)
Compliance: AWS lists most of these as HIPAA eligible, but the developer must confirm the current list and sign the AWS Business Associate Addendum if patient data is involved.
Phases
Five phases reach a live pilot in roughly 10-14 weeks (estimates to confirm with your developer).
1. Setup (1 week). Your team sends demos, FAQ and CRM choice; developer prepares AWS and Bedrock access.
2. Prototype (2-3 weeks). One demo, 10 FAQs, voice call in a browser. Gate: your test call approves voice and flow.
3. Demo library (2-3 weeks). All 30+ demos chaptered and tagged, full FAQ loaded. Gate: your team reviews tags and sample paths.
4. Integration (2-3 weeks). Widget in your software, CRM, calendar, recap email, handoff. Gate: end-to-end test.
5. Pilot (2-4 weeks). Real customers, limited volume, human review of every call. Gate: go or no-go.
Later, optional: agent operating a live copy of your software, phone calls, avatar, more languages.
Risks and open decisions
Six decisions from your team unlock the build.
[ ] Viewer: a buyer watching a demo, or a patient being guided through the software?
[ ] Which CRM and calendar?
[ ] Which languages at launch?
[ ] Does any demo show real patient data?
[ ] Avatar, or voice plus shared demo screen only?
[ ] Who receives handoffs, and how fast must they respond?
Risks
• Voice delay: replies slower than about one second feel unnatural; stream speech and test in the prototype.
• Wrong answers: the agent answers only from the FAQ and transcripts, uses Guardrails, and flags a person when unsure.
• Patient data: needs AWS BAA, encryption, access logs and review by a qualified person; Claude can build to the requirements but cannot certify them.
• Recording consent: your legal team approves the notice wording.
• Cost per call: video, speech and Claude bill per use; measure in the prototype.