buzy-ai-gemma4-lora

Buzy AI – Business Reasoning LoRA for Gemma 4

Buzy AI is a LoRA adapter built on Gemma 4 that specializes the model for explainable business reasoning over enterprise documents.

Rather than acting as a generic chatbot, Buzy AI transforms business information into structured, evidence-based recommendations to support decision-making.

Model Details

• Developed by: Destin Biringanine
• Entreprise: Code, Growth Alive
• Base model: unsloth/gemma-4-E2B-it-unsloth-bnb-4bit
• Fine-tuning: LoRA (PEFT)
• Framework: Unsloth
• License: Apache-2.0

Intended Use

This model is designed for enterprise decision-support tasks, including:

• Contract analysis
• Invoice understanding
• Procurement intelligence
• Supplier risk assessment
• Financial document analysis
• Business reporting
• Explainable recommendations

The model follows a structured reasoning process consisting of:

• Evidence
• Reasoning
• Confidence
• Business Impact
• Recommended Actions

Training

The adapter was fine-tuned using Unsloth on a structured business reasoning dataset.

The objective was not to teach new business knowledge, but to improve the model's ability to produce consistent, transparent, and explainable business reasoning.

Example

Input

Supplier A has delayed deliveries for three consecutive months.
The defect rate increased by 18%.
The contract expires next month.
Should the contract be renewed?

Output

{
  "decision": "Do not renew without corrective actions.",
  "evidence": [
    "Three consecutive delivery delays",
    "18% increase in defect rate"
  ],
  "reasoning": "Operational performance has consistently deteriorated, increasing supply-chain risk.",
  "confidence": 0.95,
  "business_impact": "High operational risk",
  "recommended_actions": [
    "Request a corrective action plan",
    "Evaluate alternative suppliers",
    "Conduct a supplier performance review"
  ]
}

Limitations

This model is intended for research and demonstration purposes.

It should be used as a decision-support assistant rather than an autonomous decision-maker and performs best when combined with an external knowledge base or retrieval system.

Acknowledgements

This project was developed for the Build with Google Gemma Hackathon using:

• Google Gemma
• Unsloth
• Hugging Face Transformers
• PEFT
• PyTorch