# THRPY RealAffect Platform: KPIs & Key Differentiators

**Last Updated:** March 2026  
**Focus:** THRPY Edge-Native Affective Computing Infrastructure  

## 🌐 The Pitch: Platform, Engine, or Persistent Affective Field Manager?

To prevent fragmenting the THRPY AI infrastructure pitch, we must clearly distinguish the roles of our proprietary technologies. We are not selling disconnected pieces; we are providing a unified **Edge-Native Affective Computing Platform**.

Here is the exact vernacular and layered architecture:

### Layer 1 — RealFlow Runtime (The Transport)
The ultra-fast, bidirectional stream that orchestrates Audio, WebSockets, STT, TTS, and LLM turn-taking.

### Layer 2 — RealFuel Sensors (The Signals)
The raw multi-modal ingestion layer analyzing Voice, Text, and Face (future).

### Layer 3 — RealFeel Core Engine (The Math)
The proprietary affective mathematics calculating Valence, Arousal, Undercurrent, and **RealBalance**. *(RealBalance stabilizes emotional trajectory modeling to prevent oscillation or false spikes).*

### Layer 4 — MMRY Fusion & State Field (The Memory Synthesis)
MMRY creates a local emotional memory graph that never leaves the user's device. Through **MMRY Fusion**, it synthesizes the immediate RealFuel signals and RealFeel math with the user's long-term vector embeddings and emotional trajectories.

### Layer 5 — RealAction Layer (The Output)
The UI and behavioral outputs. It orchestrates dynamic tool triggering, clinical risk gating (**RealSafety**), UI color shifts, and dynamic speech modulation.

## 🔄 The Persistent Affective Field Dynamics

In neuroscience terms, we model emotion as a dynamical system (measuring emotional motion and momentum, not just static classification). This infrastructure operates as a continuous4. **MMRY Fusion & State Generation:** The system synthesizes the immediate data with historical vectors, producing two concrete outputs concurrently:

- **The Conscious State:** The reported emotional state (Current Valence/Arousal).
- **The Subconscious State:** The trajectory memory and affective momentum.

---

## 🎯 Key Performance Indicators (KPIs)

Our metrics reflect a radical departure from cloud-dependent architectures, achieving speeds previously impossible by moving inference directly to the browser edge.

### Latency Metrics (The Edge Advantage)

| Metric | Target | Actual | Industry Standard | Advantage |
|--------|--------|--------|-------------------|-----------|
| **Vocal Extraction (Edge)** | <30ms | **~22ms** | 200-500ms (Cloud) | **10-20x faster** (Web Audio Worklet directly processes ZCR/Pitch) |
| **STT Latency** | <300ms | 100-300ms | 500-1000ms | **2-3x faster** |
| **Text Inference** | <300ms | 100-300ms | 300-800ms | **2-3x faster** |
| **Local Embeddings (MMRY)** | <100ms | <50ms | 300-600ms (API) | **Instant & 100% Private** |
| **Safety Check** | <50ms | 20-50ms | 100-300ms | **2-6x faster** |
| **LLM First Token** | <200ms | 200-1000ms | 500-2000ms | **2-10x faster** |
| **TTS Generation** | <200ms | 50-200ms | 200-1000ms | **4-20x faster** |
| **End-to-End** | <2000ms | 500-2000ms | 2000-5000ms | **2-5x faster** |

---

## 🚀 Key Differentiators (The RealAffect Edge)

### 1. Edge-Native Vocal Emotion (22ms Latency)

**What It Is:**

- We do not send audio to the cloud to understand emotion. Our Web Audio Worklet extracts Zero-Crossing Rate (ZCR) and pitch in real-time (~22ms) on the user's local device.
- These raw signals are fed into the `@thrpy/core-engine` to map mathematical vocal metrics.

**Why It Matters & Competitive Advantage:**

- Sending raw audio arrays to the cloud for inference incurs massive I/O latency (often 500ms+).
- THRPY's edge-native execution reduces this layer to an imperceptible **22ms**.
- It also guarantees absolute privacy since the raw acoustic traits never leave the browser.

### 2. MMRY: The Edge-Native Vector State Field

**What It Is:**

- **MMRY creates a local emotional memory graph that never leaves the user's device.**
- It securely generates 384-dimensional contextual vectors and handles abstract session summarizations using WebAssembly embeddings on the client's CPU/GPU.

**Why It Matters & Competitive Advantage:**

- **Zero API Costs & Absolute Data Sovereignty:** Competitors rely on OpenAI or Pinecone to store and generate embeddings, exposing clinical data. MMRY keeps the persistent affective field manager 100% local. It ensures therapeutic interactions are fully compartmentalized on the user's device.

### 3. Safety-Gated, RealFlow Pipeline

**What It Is:**

- RealFlow isn't just a WebSocket; it's a safety-gated bidirectional river.
- Safety gating runs locally and proxy-side **before** an LLM responds. Gate assessments take <50ms.

**Why It Matters & Competitive Advantage:**

- Standard LLM safety systems are reactive (post-processing). RealFlow's proactive gating is crucial for clinical applications where a delayed safety net could trigger a detrimental user experience.

### 4. Parallel Cognitive Processing

**What It Is:**

- The system evaluates "What they are saying" (RealFlow STT) and "How they are saying it" (Core Engine Vocal Metrics) simultaneously.
- The state field (MMRY) is continuously updated in parallel, not sequentially.

**Why It Matters & Competitive Advantage:**

- **Traditional Pipeline (Serial Processing):** `Listen -> Transcribe -> API Call -> AI Thought -> API Call -> Generate Speech -> Speak`
- **THRPY Pipeline (Parallel Inference):** We evaluate *speech tone*, *speech meaning*, *memory update*, and *risk evaluation* **all simultaneously.**

---

## 📊 Market Position: The Missing Layer of AI Infrastructure

### Where RealFlow & MMRY Fit

**Tier 1: Generic Voice AI (Dialogflow, Lex)**

- Simple intent matching, no emotional context.

**Tier 2: API-Based Voice Agents (Vapi, Bland)**

- Great latency, but fully dependent on cloud APIs, zero clinical safety gating, and zero edge-native privacy.

**Tier 3.5: Emotion APIs (Hume, Affectiva)**

- Cloud-dependent classification APIs measuring facial or vocal sentiment without providing the full affective operational runtime or privacy.

**Tier 4: The THRPY RealAffect Runtime (Us)**

- We provide the **affective runtime**—the full Emotional Operating System for AI.
- A foundational persistent affective field manager that utilizes **MMRY Fusion** to bind transport (RealFlow), signals (RealFuel), and memory into one secure edge-native layer.

### Core Value Proposition (To Investors and Developers)

**Today's Pitch:**
> *"THRPY is the first edge-native affective computing platform. We don't just route chat completions faster; we run the emotional intelligence natively in the browser. By extracting vocal metrics in 22 milliseconds on the edge, mapping it in our proprietary math engine, and organizing it in a local WebAssembly vector state field (MMRY), we provide a deeply personal, zero-latency, and 100% private therapeutic AI experience."*

**The Future Evolution:**
> *"The THRPY RealAffect Platform will become the **foundational reference runtime** for all affective computing. It acts as the emotional operating system for AI. Developers will be able to plug the Core Engine and MMRY vector fields into their own pipelines—whether in software, robotics, or cars—creating dynamic agents that remember and feel without ever compromising user data sovereignty."*