DEV Community

Cheetu AI
Cheetu AI

Posted on

Building an AI Meeting Assistant for Global Teams: What Worked, What Didn't, and What I'd Do Differently

Most AI meeting assistants promise the same thing:

"We'll record your meeting and generate a summary."

After testing several tools and building our own workflow, I realized that wasn't the problem we were trying to solve.

The real challenge wasn't creating transcripts.

It was helping distributed teams turn conversations into actions.

So instead of looking for another meeting recorder, we designed a workflow that automatically captures discussions, extracts decisions, generates summaries, translates key information, and shares everything with the right people.

Here's what we learned.


Why We Started

Our team works across multiple countries.

That means almost every week includes:

  • Product planning meetings
  • Customer interviews
  • Internal engineering discussions
  • Cross-functional project reviews

Each meeting generated the same repetitive work:

  • Someone wrote notes.
  • Someone summarized them.
  • Someone translated them.
  • Someone copied action items into project management tools.
  • Someone answered questions from teammates who couldn't attend.

None of this work created new value.

It was simply necessary administration.

We wanted to automate it.


Designing the Workflow

Rather than depending on a single AI model, we split the workflow into several independent stages.

Meeting Audio
      │
      ▼
Speech Recognition
      │
      ▼
Transcript Cleanup
      │
      ▼
LLM Analysis
      │
      ├─────────────► Key Decisions
      ├─────────────► Action Items
      ├─────────────► Meeting Summary
      └─────────────► Open Questions
                     │
                     ▼
          Multilingual Translation
                     │
                     ▼
      Slack • Email • Knowledge Base
Enter fullscreen mode Exit fullscreen mode

Each stage has one job.

That makes the system easier to maintain and significantly easier to improve.


Lesson #1: Transcripts Are Just Raw Data

At first, we were excited about accurate speech recognition.

Then we discovered something interesting.

Nobody wanted to read a 9,000-word transcript.

What everyone actually wanted was:

  • What was decided?
  • What should I do?
  • What changed?
  • What still needs discussion?

The transcript became an input—not the final product.


Lesson #2: Prompt Engineering Matters More Than We Expected

Initially, our prompt looked something like this:

Summarize this meeting.

The results were inconsistent.

Sometimes it focused on introductions.

Sometimes it ignored important decisions.

Sometimes it produced summaries that looked impressive but weren't actually useful.

We replaced that with structured instructions.

Instead of asking for "a summary," we asked the model to produce specific sections:

  • Executive Summary
  • Key Decisions
  • Action Items
  • Risks
  • Open Questions
  • Next Steps

The quality improved immediately.

Giving the model a framework produced much more predictable output.


Lesson #3: Translation Should Happen Last

This one surprised us.

Our first version translated the entire transcript.

It worked.

It was also expensive, slow, and unnecessary.

Most people don't need a translated transcript.

They need translated information.

Now we only translate:

  • Meeting summaries
  • Decisions
  • Action items

The output is smaller, faster, and much easier to review.

For multilingual teams, this simple change made a noticeable difference.


Lesson #4: Automation Should Feel Invisible

The best automation isn't something people notice.

It's something they stop thinking about.

Our goal wasn't to build another dashboard.

It was to make meeting follow-up happen automatically.

After each meeting:

  • summaries are generated
  • action items are extracted
  • multilingual versions are created
  • results are shared with the team

Nobody has to remember the process.

The workflow simply continues after the meeting ends.


Where Cheetu AI Fits

One of the remaining challenges was multilingual collaboration.

Many AI tools generate excellent English summaries but require additional manual work when teams operate across multiple languages.

We introduced Cheetu AI as the multilingual layer in our workflow.

Instead of translating raw transcripts, it helps convert structured meeting outputs into multilingual summaries that are easier to distribute across global teams.

By placing translation after summarization, we reduced unnecessary processing while maintaining context and consistency.


What Changed?

The biggest improvement wasn't speed.

It was alignment.

Instead of everyone creating their own version of the meeting, the team worked from one structured output.

That reduced confusion, shortened follow-up discussions, and made meetings easier to search later.

More importantly, engineers could focus on solving problems instead of documenting conversations.


If I Were Building This Again...

I'd follow the same principles:

  1. Treat transcripts as input, not output.
  2. Design prompts around business outcomes instead of generic summaries.
  3. Translate structured information instead of raw conversations.
  4. Build workflows, not isolated AI features.
  5. Make automation invisible to the end user.

The technology will continue to improve.

But I think the biggest opportunity isn't better transcription.

It's building systems that turn conversations into knowledge—and knowledge into action.

That's where AI starts creating real value.


I'd Love to Hear Your Approach

If your team is experimenting with AI for meetings, documentation, or multilingual collaboration:

  • How are you handling meeting summaries?
  • Are you translating transcripts or structured summaries?
  • What part of your workflow still requires manual effort?

I'm always interested in seeing how other developers are approaching this problem.

Top comments (0)