Decisio

AI-powered meeting decision tracker. Paste a transcript or record audio — Decisio extracts every decision made, links it to past related decisions, tracks accountability across meetings, and surfaces patterns through ML.

Features

• Decision extraction — GPT-4o-mini parses transcripts into structured decisions (statement, context, confidence, related actions, assignee).
• Voice input — record meetings in-browser; Whisper transcribes, then GPT extracts.
• Cross-meeting lineage — every new decision is matched against unresolved past ones. Status transitions: New → Updated → Conflicted → Resolved. Resolved decisions exit the duplicate-detection pool.
• Topic clustering — TF-IDF + K-Means groups decisions by theme.
• Priority prediction — Random Forest classifies decisions as High / Medium / Low. Users can override the model's prediction; the RF retrains on those labels.
• Decision timeline — visual lineage view showing how a decision evolved across meetings.

Stack

Layer	Tech
Frontend	React 18 (CRA), Framer Motion, Lucide icons
Backend	FastAPI, SQLAlchemy, SQLite
AI	OpenAI gpt-4o-mini (extraction), whisper-1 (audio)
ML	scikit-learn (TF-IDF, K-Means, Random Forest)

Setup

Prerequisites

• Python 3.10+
• Node.js 18+
• OpenAI API key

Backend

cd backend
pip install -r requirements.txt
echo "OPENAI_API_KEY=sk-..." > .env
python -m uvicorn main:app --reload --port 8001

Frontend

cd frontend
npm install
npm start

Open http://localhost:3000 — the API runs at http://localhost:8001 (docs at /docs).

API

Method	Endpoint	Purpose
POST	/analyze	Extract decisions from text
POST	/analyze-audio	Transcribe + extract from audio
GET	/meetings	List all meetings
GET	/meetings/{id}	Get one meeting + decisions
DELETE	/meetings/{id}	Delete meeting
GET	/decisions/{id}/lineage	Full parent/child decision chain
PATCH	/decisions/{id}/status	Set New / Updated / Conflicted / Resolved
PATCH	/decisions/{id}/priority	Set High / Medium / Low (or null to clear)
POST	/ml/cluster	Run topic clustering
POST	/ml/train-priority	Train RF on existing decisions
POST	/ml/predict-priorities	Predict + persist priorities
POST	/ml/insights	Combined clustering + priority output

How priority prediction works

1. Heuristic baseline — keyword lists (urgent, critical, blocker, …) score each decision; metadata signals (action count, model confidence) add boosts.
2. Random Forest — TF-IDF text features + 9 hand-crafted metadata features (length, word count, punctuation, negation, keyword counts, etc.).
3. User override — when you set a priority via the UI, that label takes precedence during training. Roughly 20–30 user labels and the model starts adapting beyond the keyword rules.

Project structure

backend/
  main.py          # FastAPI routes
  processor.py     # OpenAI extraction + prompt building
  database.py      # SQLAlchemy models + lightweight migrations
  schemas.py       # Pydantic request/response models
  ml_models.py     # TopicClusterer + PriorityPredictor
frontend/
  src/
    App.js
    components/    # DecisionCard, DecisionTimeline, VoiceRecorder, HistoryView, InsightsView
    services/API.js

License

MIT