MindMapper

AI-powered therapy session summarizer — upload a therapy transcript and get a structured SOAP note plus an interactive emotion timeline chart.

The Problem

Therapists spend significant time writing clinical notes after sessions. MindMapper automates this by generating structured SOAP (Subjective, Objective, Assessment, Plan) notes and visualizing emotional patterns across a session — all running locally with no data sent to external APIs.

Architecture

┌─────────────────────────────────────────────────────┐
│                   Gradio UI (ui.py)                  │
│  ┌──────────────┐  ┌────────────┐  ┌──────────────┐ │
│  │ Transcript   │  │ SOAP Note  │  │  Emotion     │ │
│  │ Input        │  │ (Markdown) │  │  Timeline    │ │
│  │ (.txt/paste) │  │            │  │  (Plotly)    │ │
│  └──────┬───────┘  └─────▲──────┘  └──────▲───────┘ │
└─────────┼────────────────┼─────────────────┼─────────┘
          │                │                 │
          ▼                │                 │
┌─────────────────┐  ┌─────┴───────┐  ┌─────┴────────┐
│  model.py       │  │  soap.py    │  │ emotions.py  │
│  Mistral-7B     │◄─┤  SOAP       │  │ Emotion      │
│  4-bit QLoRA    │◄─┤  Generation │  │ Extraction   │
│  + LoRA adapter │  │  + Parsing  │  │ + Plotly     │
└─────────────────┘  └─────────────┘  └──────────────┘
          ▲
          │ (optional fine-tune)
┌─────────┴─────────────────────────┐
│  training/                         │
│  train.py + dataset.py + config   │
│  QLoRA fine-tuning pipeline       │
└───────────────────────────────────┘

Setup

Local (requires NVIDIA GPU with 8GB+ VRAM)

git clone https://github.com/NotArnav03/mindmapper.git
cd mindmapper
pip install -r requirements.txt
python -m app.ui

Google Colab (recommended — free T4 GPU)

See notebooks/COLAB_GUIDE.md for step-by-step instructions.

Usage

Run the Gradio App

python -m app.ui

This launches a web interface where you can:

1. Paste a therapy transcript or upload a .txt file
2. Click "Analyze Session" to generate a SOAP note and emotion timeline
3. View the structured SOAP note in markdown format
4. Explore the interactive emotion timeline chart
5. Access raw SOAP data in JSON format

Fine-tune on Therapy Data

python training/train.py

This runs QLoRA fine-tuning with the config in training/config.yaml. The adapter is saved to ./lora_adapter and automatically loaded on next model start.

T4 GPU Memory Usage

Operation	VRAM Usage	Time
Model load (4-bit)	~5 GB	3-4 min
SOAP inference	~7 GB peak	15-30 sec
Emotion extraction (5 chunks)	~7 GB peak	60-90 sec
Fine-tuning (batch=2, grad_accum=4)	~12 GB peak	~45 min

Example Output

SOAP Note

## Subjective
- Patient reports intense work stress with multiple overlapping deadlines
- Describes crying episode in parking lot after working until midnight
- Reports impostor syndrome feelings that intensify with sleep deprivation
- Acknowledges difficulty relaxing and guilt around non-productive activities

## Objective
- Patient appears fatigued but engaged in session
- Shows emotional awareness when connecting sleep patterns to negative self-talk
- Demonstrates willingness to try behavioral interventions

## Assessment
- Work-related anxiety with significant impact on sleep quality
- Impostor syndrome exacerbated by chronic sleep deprivation
- Self-worth heavily tied to productivity — pattern recognized by patient
- Notable progress in self-awareness compared to 6 months ago

## Plan
- Replace phone-checking with 4-7-8 breathing on 2 nights this week
- Observe morning self-talk patterns following breathing exercises
- Follow-up next Thursday to review sleep intervention outcomes
- Consider structured relaxation techniques in future sessions

Tech Stack

• Model: Mistral-7B-Instruct-v0.2 with 4-bit quantization
• Fine-tuning: QLoRA via PEFT + bitsandbytes
• Training: TRL SFTTrainer
• UI: Gradio
• Visualization: Plotly
• Data: HuggingFace Datasets