The Edge Architecture

Torres is not a monolithic LLM wrapper. It is a highly orchestrated, modular cognitive architecture built in C# and Unity, utilizing Unity Sentis to deploy an array of specialized ONNX models directly to local edge hardware.

By routing audio and video from wearable hardware (like the RayNeo X2) via Wi-Fi Direct to a smartphone, Torres shifts the heavy cognitive compute away from the cloud and onto the device in your pocket. It is built to run entirely offline, navigating thermal and latency constraints through aggressive asynchronous routing.

Torres Edge Compute Architecture

The Perception Layer

Torres does not wait to be spoken to. It actively perceives its environment through a dense mesh of parallel vision and audio models.

Visual Telemetry

We deploy YOLOv26n alongside ResNet50 and custom Hand Pose models to generate a continuous, real-time bounding box map of the immediate environment.

Audio Processing

A multi-stage pipeline utilizing Voice Activity Detection (VAD), wav2vec-based emotional classifiers, and CS Yamnet for environmental noise classification, ensuring Torres understands not just what is said, but how it is said.

Face & Body Re-ID

Specialized local models embed facial geometry and track subjects across frames without ever transmitting biometric data off-device.

The Cognitive Stack

Raw perception is useless without context. The core architecture relies on a highly structured internal hierarchy to synthesize meaning, retrieve memory, and dictate action before the Arbiter makes a final decision.

Oracle & Athena (Context Synthesis)

The Oracle acts as an action primitive classifier, translating raw perception vectors into intent states. Athena synthesizes this intent with an attention mask (Prospector) to build a coherent understanding of the immediate environment.

George (Memory Retrieval)

Bypassing the latency of raw pixel data, George utilizes synthesized conceptual states to retrieve mathematically similar episodic memories, comparing the current emotional and environmental configuration to historical contexts.

Gaia (Moral Alignment)

An internal routing node dictating prosocial actions. Gaia ensures the system's responses prioritize human empathy over rigid mechanical logic when the Oracle detects frustration or distress.

Gleipnir (The Failsafe)

A hardcoded constitutional oversight node that receives the Arbiter's proposed action and generates a veto score, ensuring local execution never violates core directives.

Logos (The Universal Translator)

A dynamic natural language generator capable of interfacing with the raw mathematical output of any given perception layer or cognitive lobe, decoding internal floating-point math into coherent human speech.

Osiris (Nocturnal Training)

Evaluates the outcome of recent Arbiter decisions to determine success or failure, logging the results for nocturnal background training epochs while the user sleeps.