Swarm Intelligence System

A self-improving multi-agent ecosystem that evaluates, mutates, and selectively fine-tunes specialised agents so complex workflows become more accurate and autonomous over time.

Sustainable AI = (Owned AI × Engineering Depth × Air-Gap Integrity × Agnostic Coherence) ÷ (Platform Dependence × External Coupling) → ∞

Architecture, Design & Development by Franz Ayestarán / Collaborative Pair Programming with GitHub Copilot in VS Code

Gate Config

Window size

f_threshold

f_floor

Phi threshold

Diversity gate Require Phi >= threshold before promotion

alpha (mean fitness)

beta (min fitness)

gamma (phi)

delta (cost)

epsilon (exploration)

zeta (output entropy)

Model Endpoint Config

Model source

LLM endpoint

Model name

Timeout seconds

Training base model (HF id, optional)

Local PyTorch fallback Allow CPU fallback when Apple GPU (MPS) cannot load model

Hugging Face Sync

Hugging Face repo (owner/name)

Model alias (optional)

LoRA path (optional)

Quantization preference If unavailable, sync falls back to the best available GGUF variant.

Idle 0%

LLM Eval and Record

Temperature

Prompt

Expected answer (used for scoring)

Idle 0%

Record Fitness

Swarm id

Agent id Pick an agent from Agent Registry using Select, then run Clear Adapter.

Fitness score (0..1)

Trajectory JSON (optional)

Current Status

Decision: n/a

Samples: 0

Mean: n/a

Min: n/a

Phi: n/a

H: n/a

Niches: 0

U_SIS: n/a

Cost C: n/a

Explore E: n/a

Out Entropy: n/a

Diversity Gate: OFF

Adapter: n/a

Runtime: n/a

Ready

History

#	Score	Trajectory

Agent Registry

Readme

Adapter path (LoRA/QLoRA)

Adapter type

Training profile

Idle 0%

Swarm	Agent	Adapter	Samples	Latest	Mean	Actions