In a day packed with back-to-back discussions, preparing for meetings often gets pushed to the background. Reading long email threads, reviewing past agendas, and formulating critical discussion points manually consumes time that could be spent on actual decision-making.
To solve this, I built AI-Powered Meeting Prep Assistant. It ingests raw meeting notes, attendee lists, or brief agenda snippets and automatically generates structured executive briefings, key talking points, potential risk areas, and tailored questions—in under 5 seconds.
Here is a full breakdown of what I built, why, how it operates on AWS Free Tier services, and the main hurdles that broke during development.
đź’ˇ What & Why
The Challenge
When entering a project sync or a stakeholder call, quick context switching is tough. Traditional prep involves digging through scattered documentation, which leads to missed context or under-prepared meetings.
The Solution
PrepPulse uses foundation models hosted on Amazon Bedrock to synthesize unstructured text into structured, actionable meeting briefings. By choosing the Amazon Nova family of models, the assistant runs lightweight, fast inference with strong instruction-following capabilities—ensuring high reliability without running up cloud costs.
🏗️ Architecture & Technical Setup
The entire solution is built on a serverless architecture designed to remain within the AWS Free Tier.
1. Serverless Compute (AWS Lambda)
A Python 3.12 AWS Lambda function acts as the backend service handler. When a user pastes meeting context or attendee info, API Gateway passes the payload to Lambda.
2. Generative AI Core (Amazon Bedrock & Amazon Nova)
Lambda utilizes boto3 to interact with Amazon Bedrock using the unified Converse API.
Amazon Nova Micro / Lite: I opted for amazon.nova-micro-v1:0 for fast text-only transformations and amazon.nova-lite-v1:0 when processing richer context or meeting documents.
đź’Ą What Broke Along the Way (Lessons Learned)
Building this project taught me several critical lessons about working with Bedrock runtime endpoints and serverless configurations:
1. The Access Denied Exception & Model Entitlements
What broke: Early invocation requests directly from the Lambda function failed with access denial errors.
The fix: Beyond granting the Lambda Execution Role the bedrock:InvokeModel IAM policy, I had to manually enable access for the Amazon Nova model family in the Amazon Bedrock Console under Model Access.
2. API Schema Mismatches
What broke: Using older Bedrock payload
3. Execution Timeouts
What broke: Long, messy meeting notes pushed inference times past Lambda’s default 3-second limit.
The fix: I extended the Lambda function timeout setting to 30 seconds and optimized system prompts to constrain output tokens, reducing latency.
🙌 Inspiration & Tagged Builders
A massive shoutout to Respected mam heir summer build on served as a great practical inspiration for integrating Bedrock cleanly inside lightweight AWS Lambda execution environments.
Loading article