Skip to content

tech_stack_info.md

Latest commit

 

History

History
65 lines (43 loc) · 3.86 KB

tech_stack_info.md

File metadata and controls

65 lines (43 loc) · 3.86 KB
Raw

Latent: Technical Architecture & Stack Overview

Latent is a privacy-first, on-device Conversation Intelligence Engine built primarily to analyze live negotiations, interviews, and pitches in real-time without ever sending audio or transcriptions to the cloud.

Core Technology Stack

  • Application Framework: React Native (TypeScript)
  • Target Platforms: Android & iOS
  • State Management & UI: React Hooks, Functional Components, Custom CSS-in-JS (React Native StyleSheets).
  • Routing: React Navigation (Stack Navigator)
  • Local Storage: AsyncStorage (for saving post-session analysis & settings offline)

On-Device AI: The RunAnywhere SDKs

The entire artificial intelligence pipeline is powered exclusively by the RunAnywhere SDK. This allows massive neural network models (LLMs, STT, TTS) to run locally on the mobile phone's CPU/GPU.

The following SDK modules are integrated into Latent:

  1. @runanywhere/core
    • The foundational bridging layer that interfaces between React Native JavaScript and the underlying C++ native environment.
  2. @runanywhere/onnx
    • The ONNX Runtime backend. This powers the Speech-to-Text (STT) engine using the Sherpa Whisper Tiny model.
    • Also acts as the execution engine for the Text-to-Speech (TTS) pipeline using the Piper TTS model.
  3. @runanywhere/llamacpp
    • The Llama.cpp backend used for running full Large Language Models (LLMs) on-device, such as Liquid LFM2 (350M) or SmolLM2 (360M), without cloud latency.

Data Transmission & Input Flow

The most critical architectural pillar of Latent is the Zero-Cloud Data Policy. Here is how data is inputted, transmitted, and processed entirely within the phone's memory.

1. Audio Acquisition (Input)

  • The application requests local microphone permissions (android.permission.RECORD_AUDIO).
  • Audio is captured via native audio modules (react-native-live-audio-stream / underlying C++ audio capture buffers).
  • Transmission: Audio data is passed directly from the microphone hardware into local RAM (Float32 buffers or temporary WAV snapshots). It never leaves the device.

2. Live Transcription Pipeline (STT)

  • The SpeechService.ts module routinely chunks the incoming audio (e.g., every 1.5 to 5 seconds).
  • These localized audio snapshots are handed to the RunAnywhere.transcribeFile() C++ bridge.
  • The Sherpa Whisper Tiny ONNX model analyzes the audio purely on local hardware. We inject a Domain Bias Prompt natively to optimize the STT context for words involving finance, startup, and negotiation (e.g., prioritizing "salary" over "celery").
  • The C++ bridge returns the recognized text string back up to the JavaScript layer.

3. Contextual Auto-Correction & Parsing

  • Upon receiving the raw text string, Latent passes it through a zero-latency, synchronous TypeScript WhisperAutoCorrector.
  • This uses Levenshtein Distance algorithms to fuzzy-match phonetic hallucinations against the active Negotiation Mode (e.g., Job Interview vs. Investor Meeting) vocabulary.

4. Intent & Counter-Strategy Generation

  • The sanitized transcript string is injected into the SessionEngine, which feeds it to the IntentClassifier.
  • The engine evaluates the sentence against an offline Universal Scoring Engine (powered by structural pattern matching, negation rules, and numeric markers).
  • If a specific tactic is detected (e.g., Authority Pressure or Budget Objection), the CounterStrategyEngine instantly surfaces actionable suggestions on the React Native UI.

5. Final Output & Offline Storage

  • When a session ends, the full dialog map and cognitive metrics are compiled.
  • Data is saved directly to local flash storage via AsyncStorage.
  • The data is then displayed on the Outcome Replay Screen for post-meeting analysis, remaining 100% private to the physical device owner.