AMEVA-Orchestrator
High-Reliability Control Plane, Lifecycle Scheduling, and Multi-Modality Component Adapter Gateway SDK for Edge & Mobile AI
Install the official package directly into your runtime:
pip install termux-ai-orchestrator
# or
npm install -g @ameva/orchestratorThe Engineering Challenge
Executing concurrent multimodal AI models (LLM, STT, TTS, Vision, Diffusion) on mobile edge devices inevitably triggers Android Low Memory Killer (LMK) SIGKILL terminations, unauthorized RPC exposure risks, and severe compute resource contention across disparate runtimes.
The Architectural Breakthrough
AMEVA-Orchestrator is a lightweight, high-reliability control plane SDK and adapter gateway managing local AI model lifecycles and external runtime components on Android Termux and edge devices. It enforces a strict 5-tier lifecycle model (COLD, WARM, HOT, ACTIVE, ERROR) with process mutex serialization to prevent Android LMK terminations, and features native distributed smartphone clustering via P2P Tailscale mesh and zero-config UDP PIN pairing. Compliant with OpenSSF and CNCF security specifications, it provides built-in automated secret redaction and a tamper-evident SHA-256 audit logging chain.
Key Capabilities & Built-in Hardening
5-Tier Model Lifecycle Management
Strict 5-tier deterministic transitions with budget enforcement, completely preventing Android LMK kills.
Disaggregated Mobile AI Clustering
Pools idle RAM across heterogeneous smartphones and computes dynamic tensor split weights in real-time.
Zero-Config UDP PIN Pairing
Instant zero-config master-worker pairing via 6-digit one-time PIN and HMAC-signed UDP beacons.
Automated Secret Redaction Engine
Detects and masks Bearer tokens, passwords, API keys, and URL credentials across logs and SSE event streams.
Tamper-Evident SHA-256 Audit Trail
Cryptographically links all control commands into a SHA-256 hash chain for provable post-audit integrity.
9-Modality Component Adapter Gateway
Orchestrates full lifecycles for 9 modality adapters including termux-aichain, stt, tts, and vision.
Supported Compute Kernels & Operations
| Subsystem Category | Operations & Kernels | Status |
|---|---|---|
| Compute Engine | WebGPU Compute Shaders (WGSL), FP16/FP32 | Production |
| Memory Subsystem | Zero-Copy Ring Buffers, Weakref GC Pooling | Production |
| Platform Runtimes | Node.js, Chromium WebGPU, Android Termux Bionic | Production |
Canonical Usage Example
import ameva_orchestrator as orch
from termux_ai_orchestrator.client import Client
# 1. Probe available RAM and calculate dynamic tensor split across phone fleet:
ram_mb = orch.probe_node_ram_mb("192.168.43.101", user="u0_a123", port=8022)
servers, ts_str = orch.resolve_auto_tensor_split(["192.168.43.101:8022", "192.168.43.102:8022"])
print(f"Cluster Auto Split: -ts {ts_str}")
# 2. Connect to local Control Plane and inspect ecosystem status:
client = Client(base_url="http://127.0.0.1:11553", token="")
status = client.status()
print(f"Active nodes: {len(status['nodes'])}, Models: {len(status['models'])}")