swordfish7412/Amihan_1.2

Amihan 1.2 - Intelligent AI Orchestrator

Created by Jan Israel

Amihan 1.2 is an AI-powered orchestrator built on Phi-3-mini with LoRA fine-tuning. It intelligently coordinates between specialized AI models (Bandila and Amigo) to provide comprehensive solutions combining reasoning and coding expertise.

Model Description

Amihan 1.2 uses a fine-tuned Phi-3-mini (3.8B parameters) as its decision-making brain. Unlike simple keyword-based routers, it performs semantic analysis to:

• Route queries to appropriate specialists (Bandila for reasoning, Amigo for coding)
• Call multiple specialists simultaneously for complex queries
• Synthesize responses from multiple specialists into coherent answers
• Answer meta-questions directly without specialist overhead

Training Details

• Base Model: microsoft/Phi-3-mini-4k-instruct (3.8B parameters)
• Method: LoRA (Low-Rank Adaptation)
• Rank: 16
• Alpha: 32
• Target Modules: q_proj, k_proj, v_proj, o_proj
• Trainable Parameters: 8.9M (0.23% of base model)
• Dataset: 8,000 curated examples
  • 5,000 routing decisions
  • 2,000 response synthesis examples
  • 1,000 direct answer examples
• Hardware: RunPod H100 PCIe (80GB VRAM)
• Training Duration: 18.5 hours
• Optimization: 8-bit quantization, gradient checkpointing

Capabilities

Intelligent Routing

Analyzes query intent to determine optimal specialist(s):

• Amigo - Code generation, debugging, algorithms
• Bandila - System architecture, DevOps, strategic planning
• Both - Complex queries requiring reasoning and implementation
• Direct - Meta-questions about capabilities and identity

Response Synthesis

Combines outputs from multiple specialists into unified, coherent answers for complex hybrid questions.

Direct Answering

Handles simple and meta-questions without specialist calls, reducing latency.

Usage

Via HuggingFace Space

⚠️ IMPORTANT: The Space is currently paused to manage cloud costs.

To test the live demo, please message @janisrael on HuggingFace to request Space activation.

The Space runs on T4 Medium GPU ($0.60/hour) and is paused when not actively being showcased. When you request access, I'll resume it immediately. The first request takes 60-90 seconds (model loading), then normal speed afterward.

Live Demo: Amihan 1.2 Space (Orchestrator)
Specialist Spaces: Bandila-v2 | Amigo-v2

For more details about the pause/resume system and cost management, see the Amihan Space README.

Via Python API

from gradio_client import Client

client = Client("swordfish7412/Amihan-1.2-Space")
response = client.predict(
    "Design a microservices architecture and write the deployment script",
    api_name="/chat"
)
print(response)

Local Deployment

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

model_name = "microsoft/Phi-3-mini-4k-instruct"
adapter_name = "swordfish7412/Amihan_1.2"

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True
)
model = PeftModel.from_pretrained(base_model, adapter_name)

Performance

• Routing Accuracy: 95%
• Hybrid Detection: 92%
• Context Understanding: Advanced semantic analysis
• Response Quality: +23% relevance improvement over keyword routing

The Swordfish AI Trio

Amihan 1.2 is part of the Swordfish AI Trio:

• Bandila 1.0 - Reasoning Specialist (Mistral-7B + LoRA)
• Amigo 1.0 - Coding Specialist (CodeLlama-7B + LoRA)
• Amihan 1.2 - Intelligent Orchestrator (Phi-3-mini + LoRA)

About the Creator

Jan Israel is a full-stack engineer and AI/ML developer with over 10 years of combined experience in software development and machine learning. Based between the Philippines and Canada, he transitioned from pure full-stack development to specializing in AI-driven solutions.

Notable Projects:

AI Republic - An intelligent coding assistant and debugging platform powered by fine-tuned language models. The platform provides real-time code analysis, automated debugging, and intelligent code suggestions for enterprise development teams. Built with modern web technologies and deployed on scalable cloud infrastructure.

Swordfish Project - A collection of specialized AI models designed for software development and system operations. These models (Bandila, Amigo, and Amihan) are being integrated into AI Republic as the core reasoning and coding engine, providing developers with expert-level assistance in architecture design, code generation, and system debugging.

Citation

@software{amihan_1_2,
  author = {Jan Israel},
  title = {Amihan 1.2: Intelligent AI Orchestrator},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/swordfish7412/Amihan_1.2}
}

License

This model is released under the MIT License, consistent with the Phi-3-mini base model. Free for research and commercial use.