Cognitive Communication Coach: AI-Powered Conversation Reflection System on AWS
The project is a bilingual AI coaching system that analyzes a short conversation recording or transcript after the conversation ends. It produces a structured report with a summary, topic intent, weak reasoning points, improved response suggestions, and practice questions.
The system is not designed to secretly answer for users during a live conversation. Instead, it helps users review their own communication, understand where their reasoning was unclear, and practice better responses for future situations.
Students, interns, and junior professionals often struggle in conversations even when they understand part of the topic. Common issues include:
The problem is not only language fluency. It is also reasoning structure, confidence, and reflection.
| Service | Role | Reason for Selection |
|---|---|---|
| Amazon S3 | Stores uploaded audio, transcripts, and generated reports | Durable, low-cost object storage with private access controls |
| AWS Lambda | Handles upload URL creation, job creation, and result retrieval | Serverless compute suitable for short backend tasks |
| Amazon API Gateway | Exposes REST API endpoints for the frontend or test client | Managed API layer with authentication and throttling options |
| AWS Step Functions | Orchestrates transcription and AI analysis workflow | Makes async processing visible, retryable, and easier to debug |
| Amazon Transcribe | Converts audio conversation files into text | Managed speech-to-text service, avoids building custom ASR |
| Amazon Bedrock | Generates structured coaching feedback | Managed foundation model access without managing model infrastructure |
| Amazon DynamoDB | Stores job metadata, status, and report references | Serverless NoSQL database with simple key-value access pattern |
| Amazon CloudWatch | Stores logs and operational metrics | Required for debugging, monitoring, and validation |
| AWS IAM | Controls service permissions | Supports least-privilege security design |
| Phase | Duration | Deliverables |
|---|---|---|
| Requirement analysis | Week 1 | Problem statement, user journey, project scope |
| AWS foundation | Weeks 2-3 | IAM review, S3 bucket, DynamoDB table, basic Lambda |
| AI workflow | Weeks 4-6 | Transcribe flow, Bedrock prompt, Step Functions workflow |
| API and validation | Weeks 7-8 | API Gateway endpoints, sample input testing |
| Documentation | Weeks 9-10 | Workshop steps, diagrams, screenshots, bilingual content |
| Optimization and defense | Weeks 11-12 | Security review, cost cleanup, final rehearsal |
This project is designed for low bootcamp-scale usage:
The final report should include a screenshot or exported estimate from AWS Pricing Calculator after the student’s actual region and model choice are confirmed.
| Risk | Impact | Mitigation |
|---|---|---|
| Audio transcription is inaccurate | AI report quality decreases | Use clear sample audio, allow transcript upload fallback |
| Bedrock output is too vague | Weak demo value | Use a structured prompt and fixed evaluation rubric |
| Cost grows from repeated AI calls | Unexpected charges | Limit file duration, delete test files, monitor usage |
| Privacy concerns | Sensitive conversation data exposure | Private S3, least-privilege IAM, retention cleanup, consent notice |
| Workflow errors are hard to debug | Implementation delays | Use Step Functions execution history and CloudWatch logs |
| Scope creep into real-time assistant | Project becomes unrealistic | Keep MVP async and post-conversation only |