VORTEX

Transform Conversations into Actionable Insights Instantly

Built with the tools and technologies:

Vortex is a *Meeting Transcript Analysis * powered by *Streamlit, **NLP pipelines, and a *hybrid AI agent. It allows you to upload a meeting transcript (JSON format) and explore insights such as:

• Speaker contributions (who talked the most, speaking turns, time distribution).
• Keyword analysis (frequency, mentions, timelines, word clouds).
• Sentiment (line charts, heatmaps, distributions).
• Conversation dynamics (flow graphs, interruptions, networks).

The system is designed as a hybrid LLM + deterministic pipeline model:

• ✅ Deterministic Path → For common meeting insights, mapped to pre-coded visualizations.
• ⚡ Generative Path → For custom or unusual queries, LLM generates code dynamically.

This ensures reliability for standard tasks but also flexibility for exploratory analysis.

---

🚀 Features

• Upload meeting transcript JSON (local MVP, no cloud storage).

• Preprocessing pipeline for:

  • Parsing transcript into DataFrame.
  • Extracting keywords, relations, NER.
  • Sentiment analysis (utterance-level + trends).
  • Conversation stats.

• Pre-built visualizations:

  • Bar, pie, line, heatmap, network, word cloud, Gantt-style timeline.

• Interactive LLM Agent:

  • Maps queries to known intents (speaker_contribution_bar, keyword_frequency, etc.).
  • Falls back to generative visualization code for new queries.

• Extendable — add new pipeline functions, visualizations, or intents easily.

---

Feature	Type of Visualization	Demo Image
Speaker Contribution	Bar Chart
Sentiment Trend	Line Chart
Conversation Flow	Network
Interaction Network	NetworkX Graph

---

📂 Project Structure

transcript-ai-agent/ │ ├── backend/ │ ├── pipeline.py # Data processing & enrichment │ ├── visualization.py # Pre-coded visualizations │ └── agent.py # Intent classification & hybrid agent │ ├── app/ │ └── main.py # Streamlit frontend │ ├── requirements.txt # Python dependencies ├── .env # API key (Gemini/OpenRouter) └── README.md # Documentation

---

⚙ Setup Instructions

1. Clone the Project

bash git clone https://github.com/yourusername/transcript-ai-agent.git cd transcript-ai-agent

2. Create Virtual Environment

bash python -m venv venv source venv/bin/activate # Mac/Linux venv\Scripts\activate # Windows

3. Install Requirements

bash pip install -r requirements.txt

4. Set Environment Variables

Create a .env file:

env

If using Gemini

GEMINI_API_KEY=your_gemini_key_here

OR if using OpenRouter

OPENROUTER_API_KEY=your_openrouter_key_here

5. Run Streamlit App

bash streamlit run app.py

---

📊 Transcript Format

The transcript must be a JSON file with utterances. Example:

json [ { "speaker": "Alice", "start_time": 0, "end_time": 5, "text": "Welcome everyone to the project kickoff meeting." }, ]

---

🏗 Architecture

🔹 Pipeline (pipeline.py)

• parse_transcript → Load JSON → DataFrame.
• enrich_transcript → Add keywords, sentiment, NER, relations.
• compute_sentiment_trends → Speaker-level mood shifts.
• compute_stats → Aggregate stats (word counts, durations, etc.).

👉 This ensures every transcript is normalized before visualization.

---

🔹 Agent (agent.py)

1. Classify intent:

   • Maps query → predefined INTENTS.
   • If no match → "generative".

2. Run deterministic path:

   • Executes visualization function directly.

3. Run generative path:

   • Uses LLM to generate new Streamlit + Plotly code dynamically.
   • Code is validated and executed in sandbox.

👉 Hybrid design = safe defaults + flexibility.

---

📊 Available Visualizations

Speaker Analysis

• Contribution (bar/pie).
• Speaking time (seconds).
• Number of speaking turns.
• Utterance length distribution.
• Questions asked per speaker.

Keywords

• Keyword frequency.
• Keyword timeline.

Sentiment & Emotions

• Sentiment trend (line).

Conversation Dynamics

• Conversation flow (Graphviz).
• Conversation timeline (Gantt).
• Speaker interaction network.

---

⚡ Tech Stack

• Frontend: Streamlit
• Visualization: Plotly, Matplotlib, Seaborn, WordCloud, Graphviz, NetworkX
• NLP: spaCy (relations, NER), sentiment analysis
• LLM API: Gemini / OpenRouter
• Backend: Python (pandas, regex, heuristics)

---

🛡 Hybrid Model Explanation

Deterministic Path

• Query → maps to fixed function.
• Always produces stable, correct charts.
• Example: "Who talked the most?" → speaker_contribution_bar.

Generative Path

• If query doesn’t match → ask LLM to write code.
• Sandbox executes custom visualization code.
• Example: "Show me a bubble chart of sentiment vs word count."

👉 This ensures MVP is both robust and flexible.

---

👨‍💻 Contributors

• Mohit Gunani Mohit Harjani Soham Labhshetwar 🎯 (Project creator)

---

📜 License

MIT License. Free to use and modify.