AWS Builder Center
Weekend Annoying Task Challenge: Smart Meeting Notes Drafter

Weekend Annoying Task Challenge: Smart Meeting Notes Drafter

Smart Meeting Notes Drafter with AWS Lambda and Amazon Bedrock

Vision & What the App Does

Every week I sit through syncs, standups, and client calls, scribbling half-sentences like "mike owns fix, eta wed" into whatever app is open. Then comes the annoying part: turning that mess into something shareable. Rewriting notes into a clean summary, pulling out who owes what by when, and drafting a polite follow-up email eats 20 to 30 minutes after every single meeting. Multiply that by five or six meetings a week and I lose hours to pure formatting work.
Smart Meeting Notes Drafter kills that chore. You paste your raw, messy notes into a simple web page, pick a tone (professional, friendly, brief, or formal), and click one button. In a few seconds you get three tabs: a clean meeting summary with attendees, key points, and decisions; a structured action-items list with owners and deadlines; and a ready-to-send follow-up email in your chosen tone. Each tab has a copy-to-clipboard button, so the whole post-meeting routine drops from half an hour to about thirty seconds.

How I Built It

I wanted the simplest architecture that could still be a real, deployed app with a public URL, so I made one deliberate decision early: everything lives in a single AWS Lambda function. One Python file, no frameworks, no build step, no external dependencies beyond boto3 (which Lambda includes by default).
The Lambda handler does simple routing. A GET request returns the entire HTML user interface as a string, including all CSS and JavaScript inline. A POST to /draft parses the notes and tone from the request body, calls Amazon Bedrock, and returns JSON. The frontend fetches the /draft endpoint with a relative URL, so the same code works whether it sits behind a Function URL or API Gateway.
The prompt engineering took a couple of iterations. My system prompt tells the model to act as an executive assistant and respond only with JSON containing three fields: summary, actions, and email, with markdown inside each. Two challenges came up. First, the model sometimes wrapped its JSON in markdown code fences, which broke parsing, so I added a small fence-stripping step with a graceful fallback that shows raw text instead of erroring. Second, the default Lambda timeout of three seconds is far too short for a model call; bumping it to thirty seconds fixed the mysterious blank failures I saw on the first deploy.
One nice surprise: AWS recently retired the Bedrock "Model access" approval page. Serverless foundation models like Nova Lite are now auto-enabled on first invoke, so the only permission work needed was a small IAM inline policy.

AWS Services Used / Architecture Overview

  • AWS Lambda (Python 3.14): hosts the entire app, both UI and API, in one function
  • Lambda Function URL: provides the public HTTPS endpoint, with auth type NONE for open access
  • Amazon Bedrock (Nova Lite, via the Converse API): the AI that transforms notes into summary, actions, and email
  • AWS IAM: an inline policy grants the Lambda execution role bedrock:InvokeModel on the Nova Lite model only
  • Amazon CloudWatch Logs: automatic logging for debugging

The flow: the browser hits the Function URL, which triggers the Lambda. GET serves the HTML page; when the user clicks "Draft it," the page POSTs to /draft, the Lambda calls Bedrock's Converse API with the notes and tone, parses the JSON response, and returns it to the browser, which renders each section as markdown.
Browser ──HTTPS──> Lambda Function URL ──> AWS Lambda
├─ GET / → HTML UI
└─ POST /draft → Amazon Bedrock (Nova Lite) → JSON
Everything fits comfortably in the Free Tier: Lambda's one million free requests per month, and Nova Lite pricing so low that a full weekend of testing cost pennies.

What I Learned

This was my first time using Bedrock's Converse API, and I was impressed by how clean it is: system prompt, messages, inference config, done. I learned that structured JSON output from an LLM needs defensive parsing (strip code fences, always have a fallback), that Lambda Function URLs remove the need for API Gateway entirely for simple apps, and that a Lambda can happily serve a full interactive frontend, not just APIs. The biggest lesson was restraint: my first instinct was S3 for the frontend, API Gateway, maybe DynamoDB for history. Cutting all of that made the app faster to build, easier to explain, and trivial to maintain. One function, one URL, one annoying task gone.

Link to App

Try it: paste any rough meeting notes, pick a tone, and click Draft it.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article