AI Action Item Extractor: Meeting Dialogue to JSON.

Project Overview

Goal: Build a specialized AI model that reads meeting dialogues and extracts action items into a clean, structured JSON format.

Value Proposition:

Transforms unstructured meeting conversations into actionable, trackable tasks. Eliminates manual note-taking and follow-up confusion. Demonstrates advanced NLP and structured output generation skills.

This project transforms unstructured meeting dialogues into perfectly structured, actionable tasks. By fine-tuning and comparing two powerful language models (Mistral-7B and Phi-4), this system extracts tasks, owners, and deadlines, outputting them into a clean, machine-readable JSON format.

---

🔍 The Problem: The High Cost of Unproductive Meetings

Every professional has felt it: meetings that end with ambiguous takeaways and tasks that fall through the cracks. This isn’t just an annoyance; it’s a massive drain on productivity. Key action items are often lost in messy notes or verbal-only agreements, leading to confusion, missed deadlines, and wasted resources.

Our Solution: An intelligent system that listens to the conversation and provides a definitive source of truth for all action items, ready for integration into any project management workflow.

---

🚀 Live Demo: From Chaos to Action

Give the model a raw meeting transcript…

Input Dialogue:

#Person1#: Hey team, let's sync on the Q3 report.
#Person2#: I can take the lead on gathering the sales data. I'll have it done by Friday.
#Person1#: Great, thanks Alex. Maria, can you handle the marketing metrics?
#Person3#: Sure thing. I'll get those to you by end of day Thursday.

…and get a perfectly structured JSON output.

Generated Output:

{
  "action_items": [
    {
      "owner": "Alex",
      "task": "Gather sales data for Q3 report",
      "deadline": "by Friday"
    },
    {
      "owner": "Maria",
      "task": "Provide marketing metrics for Q3",
      "deadline": "end of day Thursday"
    }
  ]
}

---

🌟 Model Showdown: Mistral 7B vs. Phi-4

A head-to-head comparison reveals a clear trade-off between raw power and nimble efficiency.

Feature	🔵 Mistral-7B (Fine-Tuned)	🟢 Phi-4 (Fine-Tuned)
✅ JSON Validity	100% (Perfectly reliable)	90.0% (Excellent)
🎯F1-Score (Accuracy)	0.92 (High precision)	0.87 (Very strong)
⏱️ Avg. Latency (Colab T4 GPU)	~22.7 seconds	~4.2 seconds (5.4x faster)
💾 Adapter Size	~26 MB	~28-40 MB
🏆 Best For	Critical, high-stakes tasks	Real-time, speed-sensitive applications

Key Takeaway: Mistral-7B is the champion for accuracy and reliability. Phi-4 is the top choice for applications where speed is paramount.

---

🧠 How It Works: The Technical Architecture

This project leverages state-of-the-art techniques for efficient model customization.

1. Core Technology: QLoRA Fine-Tuning

Instead of training multi-billion parameter models from scratch, we use QLoRA (Quantized Low-Rank Adaptation).

# The key to memory efficiency
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

This method loads the base model in 4-bit precision and only trains small, injectable “adapter” layers, reducing memory footprint by over 99% and making it possible to train on a single consumer GPU.

2. Fine-Tuning Workflow

The process is simple, robust, and repeatable.

graph LR
    A[Raw Dialogue] --> B{Prompt Template};
    B --> C(Fine-Tuned LLM);
    C --> D[Generated Response];
    D --> E((Structured JSON));

---

📊 Performance Deep Dive

🔷 Mistral 7B: The Accuracy Powerhouse

Training Performance	Model Footprint
Loss consistently decreases, showing effective learning.	The adapter is a tiny 26 MB, a 99.8% size reduction.

Semantic Accuracy	Structural Integrity
Achieves a top-tier F1-Score of 0.92.	Delivers 100% valid JSON outputs.

---

🔶 Phi-4: The Speed & Efficiency Champion

Semantic Accuracy	Structural Integrity
A strong F1-Score of 0.87, slightly behind Mistral.	Excellent 90% JSON validity, with rare failures on complex cases.

Inference Latency vs. Input Size
Maintains low latency, with a predictable outlier on a noisy, conversational prompt (Case 5).

---

🛠️ Getting Started

Run the model yourself in just a few steps.

1. Prerequisites

• NVIDIA GPU with >= 8GB VRAM
• Python 3.10+
• A Hugging Face account (for downloading base models)

2. Installation

# Clone the repository
git clone https://github.com/your-username/your-repo-name.git
cd your-repo-name

# Create a virtual environment and install dependencies
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

requirements.txt

torch
transformers
bitsandbytes
accelerate
peft
trl
datasets

3. Run Inference Use the provided inference.py script. Update model_name and adapter_path to experiment with either model.

# inference.py
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch

# --- Configuration (Switch between models here) ---
model_name = "mistralai/Mistral-7B-v0.1"
adapter_path = "./adapters/mistral-7b-action-extractor-adapter" # UPDATE PATH

---

🔮 Future Vision & Roadmap

This project provides a strong foundation for a full-fledged productivity tool.

Next Steps (Q4 2024):

• 🏗️ Deploy as API: Wrap the model in a FastAPI service for easy integration.
• 🔄 Expand Schema: Add priority and status fields to the JSON output.
• 🤖 Improve Coreference: Better handling of pronouns (e.g., “I’ll do it,” “She will handle that”).

Long-Term Vision:

graph TD
    A[Live Meeting Audio] --> B{Speech-to-Text};
    B --> C[Real-time Action Item Extraction];
    C --> D{Automated Task Creation in Jira/Asana};
    D --> E[Progress Monitoring & AI-Generated Summaries];

---

🤝 How to Contribute

Contributions are welcome! Whether it’s adding new dialogue data, improving model quantization, or building new integrations, your help is valued.

1. Fork the repository.
2. Create a new feature branch.
3. Commit your changes and open a Pull Request with a detailed description.

---