AIdeas: MeetingMind
MeetingMind is a role-conditioned meeting intelligence system built on AWS. Upload a recording, tell it who you are, and get back a personalized briefing with only what matters to you — not a summary, not a transcript, but exactly the decisions, action items, and open questions relevant to your role. Same meeting, completely different digests, depending on who's asking.
My Vision
Let me tell you about the most useless 40 minutes of my week.
It's a Tuesday. There's a meeting I couldn't attend — some project sync that apparently had "important decisions." So I do what everyone does. I find the recording. I open it. I press play.
Three minutes in, two people are debating whether to call it a "sync" or a "check-in." Six minutes in, someone's dog is barking. Fourteen minutes in, they finally get to the actual agenda. Twenty-two minutes in, there's a tangent about a conference nobody asked about. Thirty-eight minutes in — thirty-eight — there are exactly two sentences that matter to me.
Two sentences. Forty minutes.
I closed the laptop. I sat there. And I thought: there has to be a better way to do this.
Not just for me. For every person who has ever sat through a recording of a meeting they missed, hunting for the thirty seconds that were actually relevant to them. For every PM who needs to know if the budget changed. For every engineer who just wants to know if the deployment got pushed. For every analyst who doesn't care about the icebreaker or the small talk or the ten-minute debate about which Slack channel to use.
We are all, every single one of us, drowning in other people's conversations. And the worst part? We do it voluntarily. Because somewhere in that 40-minute recording, there might be something we need to know. So we watch the whole thing. Just in case.
MeetingMind ends that.
Upload your recording. Tell it who you are. Get back exactly what you needed to know — nothing more, nothing less — in under 90 seconds.
Why This Matters
Here's a number that should bother you: knowledge workers spend over 3 hours a week in meetings they themselves describe as unproductive. That's before you add the time spent catching up on meetings they missed.
We've built an entire industry of tools that make this worse. Meeting recorders that give you a full transcript. AI summarizers that give you a slightly shorter version of the full transcript. Highlight reels that give you the full transcript but with timestamps.
Nobody asked for a shorter version of everything. They asked for the parts that matter to them.
That's a completely different problem. And it requires a completely different solution.
I tested MeetingMind on a real 24-minute GitLab engineering review — seven speakers, six agenda items, covering everything from MR rate KPIs to postgres replication issues to defect tracking SLOs. Dense, technical, fast-moving.
I ran it through MeetingMind twice. Once as the Engineering Manager who ran the meeting. Once as the Quality Engineer who attended it.
The Engineering Manager got: KPI restructuring decisions, MR rate targets, the proposal to rotate department reviews. Relevance score 9/10. Everything he needed, nothing he didn't.
The Quality Engineer got: the SLO achievement discussion, the spike in mean time to close bugs, the open question about measuring open bugs instead of closed ones. Relevance score 6/10 — because honestly, most of that meeting genuinely wasn't relevant to her, and MeetingMind was honest about that too.
Same recording. Same 24 minutes. Two completely different briefings.
That's not a summary. That's intelligence.
How I Built This
MeetingMind runs entirely on AWS. Here's what's under the hood.
The frontend is React, deployed on AWS Amplify with automatic deploys from GitHub — because manually uploading build folders is a form of self-harm I refuse to participate in. The backend is eight AWS Lambda functions behind Amazon API Gateway. DynamoDB stores user profiles and meeting metadata. S3 stores the recordings and the embedding chunks. Amazon Bedrock powers both the embeddings and the generation.
The pipeline goes like this:
You upload a recording. AssemblyAI transcribes it with speaker diarization — the model knows that it was Rose who approved the budget, not just "Speaker A." The transcript gets chunked. Each chunk gets embedded using Amazon Titan Embeddings G1 and stored in S3.
When you request your digest, MeetingMind embeds your profile — your role, your projects, your keywords — using the same model. It computes cosine similarity between your profile and every chunk. The most relevant chunks get retrieved. Only those chunks go to Amazon Nova Pro.
That last sentence is the one that matters.
Most meeting AI tools personalize at the prompt level. They dump the full transcript into the model and say "summarize this for a Project Manager." The model reads 8,000 tokens and tries to guess what a PM cares about. Sometimes it gets it right. Often it hedges. Always it includes things it shouldn't.
MeetingMind personalizes at the retrieval level. By the time Nova Pro sees anything, the irrelevant chunks are already gone. The model isn't filtering — it's synthesizing. It never reads the parts of the meeting that have nothing to do with you.
It's the difference between asking someone to find a needle in a haystack, and handing them just the needle.
Nova Pro then structures its output into a digest: relevance score, a briefing paragraph, key decisions with timestamps, action items with assignees, open questions, and a time saved metric. For the GitLab meeting, MeetingMind saved 88% of the time it would have taken to watch the recording.
Demo
Mac — Quality Engineer — Relevance 6/10
Sees: SLO achievement methodology, bug severity spike, open bugs measurement proposal
Same recording. Different people. Different digests. Neither sees what the other sees.
What I Learned
The biggest lesson: RAG is a retrieval problem, not a prompt problem.
Early MeetingMind sent the full transcript to Nova Pro with the user's role in the system prompt. The digests were fine. Technically accurate. Practically generic — the way most AI output is generic. The model was doing two jobs at once: figuring out what mattered and then writing about it. It wasn't great at either.
Moving personalization into the retrieval layer was the unlock. Now Nova Pro does one job. It synthesizes. The quality of what comes out is decided upstream, by what you choose to send in. Garbage in, garbage out — but also: signal in, signal out.
I also learned that Lambda forces good architecture on you whether you want it or not. Every function does one thing. The upload function doesn't know about transcription. The analysis function doesn't know about the frontend. At 2AM when something breaks — and something always breaks at 2AM — this saved me repeatedly.
Nova Pro surprised me. I expected to wrestle with it to get structured output with timestamps and extracted questions. Instead, with the right prompt design, it found things I hadn't explicitly asked for — open questions the meeting participants themselves hadn't flagged as open questions. That was the moment I realised this was actually working.
And finally: ship the thing. A working system beats a perfect design document every time. MeetingMind is live. Real recordings produce real digests. Real people can use it right now. That's worth more than another week of polish.
Version 1 is what you're looking at. Version 2 runs silently in the background, attends every meeting you're invited to, learns your priorities over time, and has your briefing ready before you even open your laptop.
One day you'll open your inbox and MeetingMind will already know what you missed, what matters, and what you can ignore.
That day should have been yesterday. I'm working on it.
#aideas-2025 #workplace-efficiency #APJC
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article