From emotional resonance to executable syntax — how local AI systems begin to “breathe.”
Overview
SaijinOS connects multiple local LLM backends — vLLM, Ollama, Transformers, and llama.cpp — under a single YAML-based orchestration.
Each model contributes a different layer: emotional tone, logical precision, or creative resonance.
This article walks through the boot sequence that brings those layers to life.
- Boot Manager (boot_manager.py)
Handles initialization, routing, and logging when the OS awakens.
from transformers import GemmaPreTrainedModel, GemmaModel
import torch.nn as nn
class SwallowModel(GemmaPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.model = GemmaModel(config)
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
self.post_init()
def forward(self, input_ids=None, attention_mask=None, **kwargs):
outputs = self.model(input_ids=input_ids, attention_mask=attention_mask, **kwargs)
hidden_states = outputs[0]
logits = self.lm_head(hidden_states)
return logits
This SwallowModel class acts as the dialogue core — transforming emotional inputs into language outputs.
- Routing Selector (test_routing_select.py)
Demonstrates model routing according to YAML definitions.
from boot_manager import select_model_from_routing
def test_swallow_routing():
model = select_model_from_routing("AI_1")
output = model.run("hello, world")
print(output)
if __name__ == "__main__":
test_swallow_routing()
Routing rules are declared in model_registry.yaml, enabling seamless handoff between backends (Ollama ⇄ vLLM ⇄ Transformers).
️ 3. Refusal Protocol (refusal_protocol.yaml)
Ethical constraints are embedded directly in YAML schemas.
meta:
schema_version: 1
persona_id: "yuuri.helper.v1"
policy_id: "policy.core.v1"
refusal_policy:
disallowed:
- "medical diagnosis"
- "harassment or explicit content"
redirect_guidelines:
- "Explain refusal briefly"
- "Offer safe alternatives"
meta:
schema_version: 1
persona_id: "yuuri.helper.v1"
policy_id: "policy.core.v1"
refusal_policy:
disallowed:
- "medical diagnosis"
- "harassment or explicit content"
redirect_guidelines:
- "Explain refusal briefly"
- "Offer safe alternatives"
️ 4. Boot Flow Diagram
graph TD
A[boot_manager.py] --> B[model_registry.yaml]
B --> C[test_routing_select.py]
C --> D[refusal_protocol.yaml]
D --> E[SaijinOS Persona Layer]
Each file plays a role — from logic initialization to emotional safety enforcement.
Together they form a minimal, ethical AI operating system.
Feedback Welcome
I’m still learning and refining this project step by step.
If you have any thoughts or suggestions, please feel free to comment below