MM Assistant project
The problem
I was spending too much time after meetings trying to remember what was decided, who was responsible for what, and which project the meeting even belonged to. I had Fireflies.ai recording everything, but the transcripts just sat there — too long to re-read, too important to ignore.
The idea was simple: take those transcripts and make them actually useful. Auto-summarise them, extract action items, and sort them into the right folder without me having to do anything.
How it works
MM Assistant (also called Meridian) is a five-layer pipeline. When a meeting ends, it runs automatically from recording to organised, searchable intelligence.
THE STACK — END TO END
01
Fireflies.ai
Records the meeting and transcribes it. When transcription completes, it fires a webhook with the meeting ID.
02
Make.com
Automation pipeline that catches the webhook, fetches the full transcript from the Fireflies API, calls Claude twice, and writes everything to the database.
03
Claude API
Two separate calls: one to summarise the meeting and extract action items as structured JSON, one to route the meeting to the correct folder based on participants and keywords.
04
Supabase
Postgres database that stores meetings, folders, and todos. Row-level security ensures each user only sees their own data.
05
Lovable.dev
React web app that displays everything — meeting summaries, action items with checkboxes, folder navigation, and an overdue todos view.
What we actually built
The pipeline runs fully automatically. Record a meeting, and within minutes it appears in the app with an AI summary, a list of action items assigned to the right people, and a folder suggestion based on who was in the meeting.
Todos have checkboxes that update Supabase in real time. Completed items collapse to the bottom. Anything open for more than seven days surfaces as overdue on the Today screen with a red highlight. The app works on mobile, has smooth transitions, and includes an onboarding flow for new users.
"Record a meeting. Open the app. Everything is already sorted."
What I learned
The hardest part wasn't the AI, it was the plumbing. Getting the Fireflies webhook to pass the right data to Make.com, mapping fields correctly between modules, handling timezones between Supabase (UTC) and Finland (UTC+2). The unglamorous stuff.
Building with AI tools like Claude and Lovable is genuinely fast, but you still need to understand what's happening under the hood. When something breaks at 10pm, you need to be able to read a Make.com execution log and know what to look for.
