---
title: AI Executive System
emoji: 🤖
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: "6.5.0"
python_version: "3.11"
app_file: app/app.py
pinned: false
---

# AI Executive System

A production-ready AI chatbot system that replicates the communication style, reasoning patterns, and personality of **Ryouken Okuni, CEO of Akatsuki AI Technologies**. The system uses a dual-LLM architecture with open-source models, designed for white-label deployment across multiple enterprise clients.

## Architecture

```
┌─────────────────────────────────────────────────────────────────────────────┐
│                           AI EXECUTIVE SYSTEM                                │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                              │
│   [User Query]                                                               │
│        │                                                                     │
│        ▼                                                                     │
│   ┌─────────────────────────────────────────────────────────────────────┐   │
│   │  LLM 1: VOICE MODEL (Llama-3.1-8B-Instruct + LoRA)                    │   │
│   │  - Fine-tuned on CEO's blog content                                  │   │
│   │  - Captures authentic reasoning patterns & communication style       │   │
│   │  - Generates CEO-style draft response                                │   │
│   └─────────────────────────────────────────────────────────────────────┘   │
│        │                                                                     │
│        ▼                                                                     │
│   ┌─────────────────────────────────────────────────────────────────────┐   │
│   │  LLM 2: REFINEMENT MODEL (Llama-3.1-8B-Instruct)                    │   │
│   │  - No fine-tuning required (prompt-based)                            │   │
│   │  - Polishes grammar, clarity, professional formatting                │   │
│   │  - Improves logical flow and argument coherence                      │   │
│   │  - Ensures cultural appropriateness for Japanese business context    │   │
│   │  - Preserves voice authenticity while improving quality              │   │
│   └─────────────────────────────────────────────────────────────────────┘   │
│        │                                                                     │
│        ▼                                                                     │
│   [Final Response to User]                                                   │
│                                                                              │
└─────────────────────────────────────────────────────────────────────────────┘
```

## Features

- **Dual-LLM Pipeline**: Voice model for authentic CEO communication + Refinement model for polish
- **QLoRA Fine-tuning**: Efficient training with 4-bit quantization
- **Japanese Business Culture Awareness**: Culturally appropriate responses
- **Hugging Face Integration**: Models stored and loaded from HF Hub
- **Gradio Interface**: Professional chat UI with custom branding
- **Comprehensive Evaluation**: Voice authenticity and factual accuracy metrics

## Project Structure

```
ai-executive/
├── data/
│   ├── raw/                  # Original blog posts (blogs.txt)
│   ├── processed/            # Cleaned and segmented content
│   └── training/             # Final JSONL training datasets
│
├── src/
│   ├── data_processing/      # Blog parsing, Q&A generation
│   ├── training/             # QLoRA/LoRA fine-tuning scripts
│   ├── inference/            # Dual-LLM pipeline
│   └── evaluation/           # Voice and accuracy metrics
│
├── app/                      # Gradio application
├── scripts/                  # CLI tools
└── notebooks/                # Jupyter notebooks for experiments
```

## Quick Start

### 1. Installation

```bash
pip install -r requirements.txt
```

### 2. Prepare Data

Place your CEO's blog content in `data/raw/blogs.txt` with the following format:

```
=== BLOG START ===
[Title of first blog post]

[Content of first blog post...]

=== BLOG END ===

=== BLOG START ===
[Title of second blog post]

[Content of second blog post...]

=== BLOG END ===
```

### 3. Process Blogs

```bash
python scripts/process_blogs.py --input data/raw/blogs.txt --output data/processed/
```

### 4. Generate Training Data

```bash
# Set your API key
export ANTHROPIC_API_KEY=your_key_here
# or
export OPENAI_API_KEY=your_key_here

# Generate Q&A pairs
python scripts/generate_training_data.py \
    --input data/processed/ \
    --output data/training/ \
    --num-pairs 500
```

### 5. Fine-tune the Model

```bash
# Run on Hugging Face infrastructure
python scripts/train_model.py \
    --dataset data/training/train.jsonl \
    --base-model Qwen/Qwen3-4B-Instruct \
    --output-repo your-username/ceo-voice-model \
    --epochs 3
```

### 6. Run the Chatbot

```bash
python app/app.py
```

Or deploy to Hugging Face Spaces:

```bash
python scripts/push_to_hub.py --space your-username/ai-executive-chatbot
```

## Configuration

### Environment Variables

Create a `.env` file in the project root:

```env
# API Keys for Q&A generation
ANTHROPIC_API_KEY=your_anthropic_key
OPENAI_API_KEY=your_openai_key

# Hugging Face
HF_TOKEN=your_huggingface_token
HF_USERNAME=your_username

# Model Configuration
VOICE_MODEL_REPO=your-username/ceo-voice-model
REFINEMENT_MODEL=meta-llama/Meta-Llama-3-8B-Instruct
```

### Training Configuration

Key hyperparameters in `src/training/train_qlora.py`:

| Parameter | Default | Description |
|-----------|---------|-------------|
| LoRA rank | 64 | Rank of LoRA matrices |
| LoRA alpha | 128 | Scaling factor |
| Learning rate | 2e-4 | Training learning rate |
| Batch size | 4 | Per-device batch size |
| Gradient accumulation | 4 | Steps before update |
| Max sequence length | 2048 | Maximum tokens per example |
| Epochs | 3-5 | Training epochs |

## Evaluation

Run the evaluation suite:

```bash
python scripts/evaluate_model.py \
    --model your-username/ceo-voice-model \
    --test-set data/training/validation.jsonl
```

### Metrics

- **Vocabulary Overlap**: Jaccard similarity with blog corpus (target: >0.7)
- **Embedding Similarity**: Topic coherence with source material (target: >0.8)
- **Factual Accuracy**: Claims verified against source (target: >95%)
- **Unique Phrase Preservation**: CEO's signature phrases detected

## Deployment

### Hugging Face Spaces

1. Create a new Space on Hugging Face
2. Push the application:

```bash
python scripts/push_to_hub.py --space your-username/ai-executive-chatbot
```

### Requirements

- GPU: NVIDIA A10G or T4 recommended
- VRAM: 16GB+ for inference
- Storage: 20GB+ for models

## License

Proprietary - Akatsuki AI Technologies

## Contact

For questions or support, contact the Akatsuki AI Technologies team.