ClarityMeet

Project Title

ClarityMeet - Meeting intelligence with transcript parsing, decisions, and minutes.

The Problem

Teams lose track of decisions and action items buried in long meeting transcripts. Manually summarizing meetings is slow and inconsistent, and it is hard to answer follow-up questions about why a decision was made.

The Solution

ClarityMeet parses uploaded transcripts, extracts decisions and action items, stores structured data and semantic embeddings, and provides a chat-style meeting intelligence endpoint. It can also generate concise meeting minutes from the stored context to make follow-up and accountability easy.

Tech Stack

• Languages: Python, JavaScript
• Backend: FastAPI
• Frontend: React (Vite)
• Database: Supabase (PostgreSQL + pgvector)
• Storage: Supabase Storage (optional)
• AI models: Cloudflare Workers AI (@cf/meta/llama-3-8b-instruct, @cf/baai/bge-large-en-v1.5)

Database Schema

extracted_items

• id (int8)
• meeting_id (text)
• filename (text)
• meeting_date (date)
• item_type (text)
• data (jsonb)
• created_at (timestamptz)

transcript_chunks

• id (int8)
• meeting_id (text)
• filename (text)
• chunk_index (int4)
• content (text)
• timestamp_start (text)
• timestamp_end (text)
• created_at (timestamptz)
• embedding (vector)

Setup Instructions

Backend

1. Create and activate a virtual environment.
2. Install dependencies:

   pip install -r requirements.txt

3. Create a .env file in the repo root with:

   SUPABASE_URL=your_supabase_url
   SUPABASE_KEY=your_supabase_key

4. Set Cloudflare Workers AI model IDs (optional overrides):

   CF_CHAT_MODEL=@cf/meta/llama-3-8b-instruct
   CF_EMBED_MODEL=@cf/baai/bge-large-en-v1.5

5. Start the API:

   uvicorn src.main:app --reload

Frontend

Hosted frontend: https://clarity-meet-five.vercel.app/

1. Install dependencies:

   cd frontend
   npm install

2. Start the dev server:

   npm run dev

Supabase Notes

• Create tables extracted_items and transcript_chunks.
• Enable pgvector and add an embedding column to transcript_chunks.
• Create an RPC function match_chunks for vector similarity search.
• Create a Storage bucket named transcripts if you want to store raw uploads.