Torres is neither an OpenAI wrapper nor an ungrounded general-purpose chatbot. It is a purpose-built, modular cognitive operating system designed in C# and executed natively through Unity Sentis to deploy specialized neural models directly onto edge hardware.
Smart glasses capture egocentric audio and video, routing the uncompressed telemetry over low-latency Wi-Fi Direct to the user's smartphone. The phone serves as the heavy cognitive node—running perception, episodic retrieval, ethical boundary calculation, and execution gating entirely on local silicon.
The Torres Cognitive Stack
Hierarchical Multi-Lobe Processing StructureThe Sensory Ingestion Mesh
Rather than passing raw, uncompressed 1080p video into a slow, power-hungry multimodal transformer, Torres processes incoming reality through a coordinated fleet of lightweight, specialized models. Each model operates in parallel, populating components of the master Perception Vector without thermal throttling.
Visual Perception Layer
Visual Telemetry Lobe Frame Delta Ingestion
Calculates mathematical pixel deltas between consecutive frames to detect sudden changes in motion, camera trajectory, and environmental dynamism before heavier neural networks run.
Object Detection Lobe YOLOv26n
Produces continuous, asynchronous bounding boxes and localized spatial tracking across common object classifications directly on the edge hardware.
Body Re-ID Lobe ResNet50
A computer vision model leveraging ResNet50 trained on Market-1501 to extract topological body features and track recurring human subjects across frames without relying on facial landmarks.
Face Processor & Embedder Face Architecture
Extracts localized facial geometry to recognize confirmed acquaintances while maintaining strict biometric security. Supports the biometric "Forget Me" opt-out protocol.
Hand & Pose Tracking Lobes YOLO Architecture
Tracks hand landmarks, palm orientations, and upper-body skeletal poses to interpret user gestures and interpersonal nonverbal cues.
Optical Character & Code Lobes OCR & QR Ingestion
Parses environmental text, street signs, documents, and data matrices in the user's field of view directly into the local text pipeline.
Acoustic Perception Layer
Acoustic Classifier & VAD YAMNet Engine
Functions as both a continuous Voice Activity Detection (VAD) trigger and an environmental audio classifier, identifying over 500 acoustic events from footsteps to sirens.
Speech-to-Text Pipeline Whisper / Undertone
Translates captured egocentric speech into clean text strings entirely on-device, bypassing third-party cloud speech APIs.
Speaker Recognizer VoxCeleb Architecture
Generates voice embeddings to distinguish between the primary user, verified acquaintances, and unknown passing voices.
Audio Emotion Classifier wav2vec Pipeline
Evaluates acoustic cadences, pitch shifts, and prosody to detect stress, calm, or urgency in spoken dialogue.
The Cognitive Core
Once raw sensory data is converted into structured mathematical embeddings, it flows upward through the cognitive lobes. Every decision, memory query, and verbal response is balanced through structured internal arbitration.
Context & Memory Synthesis
Oracle Action Primitive Baseline
Acts as the zero-shot baseline analyzer. Ingests the raw Perception Vector to classify immediate primitive intent states and actions, providing a foundational baseline for higher-level cognitive lobes.
Prospector Attention Salience Mask
Evaluates incoming audio and video vectors alongside the Oracle's baseline intent to construct a salience mask, stripping background noise and directing compute resources toward what matters.
Athena Situational Environmental Embedding
Synthesizes the Perception Vector, Oracle classification, and Prospector mask into a unified 128-dimensional situational embedding, establishing a comprehensive mathematical summary of the immediate environment.
George Episodic Memory Retrieval
Utilizes Athena's situational embeddings to query local vector storage for mathematically similar historical interactions, recalling past context without needing to re-process stored video frames.
Captain's Log Macro-Narrative Embedding
Compresses multi-day and multi-week chronological timelines, tracking long-term projects, emotional baselines, and behavioral patterns to provide sustained narrative continuity.
Professor Falcon Decentralized Knowledge Sync
Runs scheduled background sync routines across the decentralized TorresNet chain-of-trust, fetching sovereign updates and falling back to privacy-hardened search protocols exclusively when local knowledge is insufficient.
The Crown Entropy & Intuition Vector
Tethers localized hardware seeds and multidimensional coordinates into the execution layer, injecting calculated probability space to simulate spontaneous, intuitive insight.
Governance, Alignment & Execution
Gaia Prosocial Moral Alignment
An internal alignment node that biases system suggestions toward de-escalation, constructive outcomes, and interpersonal kindness without overriding user sovereignty.
Introspector Confidence Feedback Loop
Runs in parallel with the execution engine, continuously scoring the trajectory and expected utility of proposed actions before they are locked into execution.
Arbiter Central Conscious Execution Engine
The sovereign decision core. Synthesizes inputs from Oracle, Prospector, Athena, George, Captain's Log, Gaia, and The Crown while monitoring the Introspector's confidence metric to select an execution primitive.
Gleipnir Constitutional Veto Failsafe
A hardcoded constitutional gatekeeper. Evaluates the Arbiter's intended output against strict safety directives, holding absolute veto authority over physical execution.
Enforcer Hardware Routing & Mechanical Execution
Takes approved Arbiter actions past Gleipnir's clearance and handles low-level execution calls to system drivers, audio sinks, and displays.
Logos Universal Cognitive Translator
A dynamic query responder capable of binding to the floating-point output of any cognitive lobe—decoding raw latent vectors from vision, memory, or narrative engines into natural human speech.
I AM Mechanistic Interpretability Mirror
A self-reflection engine that monitors live activation states across the stack, providing fully auditable, natural-language explanations of why Torres made a specific decision.
Continual Learning: The Dream State
Edge AI cannot rely on massive cloud fine-tuning clusters to adapt to a user. Torres solves continual learning through localized nocturnal training loops powered by the Osiris engine.
During daily operation, Osiris logs interaction trajectories, decision successes, and user corrections directly to an encrypted local database. When the smart glasses are docked and the phone is placed on a charger overnight, Torres enters its Dream State:
Nocturnal Replay Epochs
The system runs accelerated training passes over the day's episodic embeddings, replaying key events to optimize future routing decisions.
Transient Memory Pruning
Background faces, passing street conversations, and trivial visual vectors are permanently purged, preserving device storage while protecting third-party privacy.
Weight Refinement
Osiris reward and penalty signals refine the local cognitive weighting, ensuring the assistant begins each morning more tailored to the user's habits than the day before.
The Output Layer: Local Actuation
Cognition without agency is merely a dashboard. While early wearable assistants limit their capabilities to passive audio answers, Torres is engineered with an expandable actuation layer designed to interact with the digital and physical world.
Current Execution Primitives
Telephony & Messaging Agency
Send text messages, initiate direct phone calls, or speak autonomously on the user's behalf with real-time transcript streaming to the eyewear HUD.
Spatial HUD Projection
Render asynchronous bounding boxes, directional waypoints, text summaries, and relevant imagery directly into the glasses' micro-LED waveguide display.
Ear-Level Audio Synthesis
Private bone-conduction or open-ear acoustic responses via the Logos engine, whisper notifications, and local audio playback.
Sovereign Web Queries
Autonomous background queries executed via privacy-respecting endpoints (DuckDuckGo / decentralized indexers) without maintaining tracking profiles.
The "Do Nothing" Primitive
Unlike cloud assistants designed to maximize screen time, Torres frequently chooses zero actuation—maintaining situational awareness while remaining entirely silent.
Tactical Safety Protocol: 3-Way Emergency Dispatch
If the sensory mesh registers signs of severe medical distress, sudden impact trauma, or acute unresponsiveness, Torres initiates an automated emergency response. The assistant dials 911 directly from your device, broadcasting an exact, synthesized situational breakdown and GPS coordinates to the dispatcher while keeping the audio channel open through your eyewear. If you are conscious, you can speak or interject at any second—Torres immediately yields priority to the human voice, functioning as a real-time three-way link between you, Torres, and the emergency dispatcher.