Building an AI-Powered Meeting Summary System (Complete Code + Architecture)

#ai #python #automation #tutorial

Meeting summaries were eating 30+ minutes per meeting. I automated the entire process — from recording to action item creation.

Architecture

Zoom/Meet Recording
        ↓
  Whisper (transcription)
        ↓
  Claude API (analysis)
        ↓
  ┌─────┼─────┐
  ↓     ↓     ↓
Email  Slack  Asana
(summary) (alert) (tasks)

Step 1: Transcription

import whisper
import os

def transcribe_meeting(audio_path):
    model = whisper.load_model("base")
    result = model.transcribe(audio_path)

    # Save transcript
    transcript_path = audio_path.replace('.mp3', '_transcript.txt')
    with open(transcript_path, 'w') as f:
        f.write(result['text'])

    return result['text']

For production, I use Whisper API or Otter.ai — faster and handles longer recordings.

Step 2: AI Analysis

import anthropic

def analyze_meeting(transcript, meeting_context=None):
    client = anthropic.Anthropic()

    context = f"Meeting context: {meeting_context}" if meeting_context else ""

    response = client.messages.create(
        model="claude-sonnet-4-20250514",
        max_tokens=2000,
        messages=[{
            "role": "user",
            "content": f"""Analyze this meeting transcript and generate a structured summary.

{context}

Transcript:
{transcript}

Generate the following sections:

## Meeting Summary
2-3 sentence executive summary of what was discussed and decided.

## Key Decisions
Numbered list of decisions made, with who made them.

## Action Items
Table format:
| Task | Owner | Deadline | Priority |
Each action item must have a specific owner and deadline.

## Open Questions
Unresolved items that need follow-up.

## Notable Quotes
Any particularly important statements (with attribution).

## Next Meeting
Suggested agenda items for the next meeting.
"""
        }]
    )
    return response.content[0].text

# Example usage
transcript = transcribe_meeting("weekly_standup_2026-03-25.mp3")
summary = analyze_meeting(transcript, "Weekly engineering standup, Q1 sprint review")

Step 3: Distribution

import smtplib
from email.mime.text import MIMEText
import requests

def distribute_summary(summary, attendees, slack_channel=None):
    # Email to all attendees
    for attendee in attendees:
        msg = MIMEText(summary, 'html')
        msg['Subject'] = f"Meeting Summary — {datetime.now().strftime('%B %d, %Y')}"
        msg['From'] = "meetings@yourdomain.com"
        msg['To'] = attendee['email']

        with smtplib.SMTP_SSL('smtp.gmail.com', 465) as server:
            server.login(EMAIL, PASSWORD)
            server.send_message(msg)

    # Slack notification
    if slack_channel:
        requests.post(SLACK_WEBHOOK, json={
            "channel": slack_channel,
            "text": f"Meeting summary posted. Action items:\n{extract_action_items(summary)}"
        })

def create_tasks(summary):
    action_items = parse_action_items(summary)

    for item in action_items:
        # Create Asana task
        requests.post("https://app.asana.com/api/1.0/tasks", 
            headers={"Authorization": f"Bearer {ASANA_TOKEN}"},
            json={
                "data": {
                    "name": item['task'],
                    "assignee": item['owner_email'],
                    "due_on": item['deadline'],
                    "projects": [PROJECT_ID],
                    "notes": f"From meeting on {datetime.now().strftime('%Y-%m-%d')}\n\n{item['context']}"
                }
            })

Step 4: Full Pipeline

def process_meeting(audio_path, attendees, meeting_type="general"):
    print(f"Processing: {audio_path}")

    # 1. Transcribe
    transcript = transcribe_meeting(audio_path)
    print(f"Transcribed: {len(transcript)} characters")

    # 2. Analyze
    context = f"{meeting_type} meeting with {len(attendees)} attendees"
    summary = analyze_meeting(transcript, context)
    print("Analysis complete")

    # 3. Distribute
    distribute_summary(summary, attendees, slack_channel="#meeting-summaries")
    print("Summary distributed")

    # 4. Create tasks
    create_tasks(summary)
    print("Action items created in Asana")

    return summary

# Automation: Watch for new recordings
from watchdog.observers import Observer
from watchdog.events import FileSystemEventHandler

class MeetingHandler(FileSystemEventHandler):
    def on_created(self, event):
        if event.src_path.endswith(('.mp3', '.mp4', '.wav')):
            # Look up attendees from calendar
            attendees = get_meeting_attendees(event.src_path)
            process_meeting(event.src_path, attendees)

Results

Metric	Before	After
Summary creation time	30-45 min	2 min (automated)
Action item tracking	"I forgot"	100% captured
Meeting follow-up rate	~60%	~95%
Time saved per week (10 meetings)	5-7.5 hrs	0 hrs (fully automated)

Production Tips

Handle audio quality: Whisper struggles with poor audio. Add noise reduction pre-processing.
Speaker identification: Use pyannote.audio for speaker diarization — critical for attributing action items.
Retry logic: API calls fail. Build retry with exponential backoff.
Privacy: Never store raw transcripts longer than needed. Process and delete.

For the complete implementation plus 29 more automation blueprints: AI Automation Playbook — $147

What's your current meeting workflow? Comment and I'll suggest the quickest automation win.

DEV Community