Latent

Privacy-first, on-device conversation intelligence for real-time tactical guidance.

Latent is a React Native app that helps you during high-stakes conversations (interviews, sales calls, investor discussions, salary negotiations) by transcribing speech locally, detecting negotiation signals, and surfacing tactical suggestions in real time.

---

What Latent does

• Captures live microphone audio
• Transcribes speech on-device
• Detects persuasion/negotiation patterns
• Shows in-session tactical counter-strategies
• Stores session history locally for replay and analysis

No cloud inference is required for core runtime behavior.

---

Current AI Runtime Stack

Latent has been migrated to an SDK-independent local inference stack:

• STT: whisper.rn (whisper.cpp)
• LLM runtime: llama.rn (llama.cpp)

Models currently wired

• Whisper STT model: ggml-base.en.bin
• LLM model: LFM2-350M-Q8_0.gguf

Notes

• Whisper context is initialized in ModelService.tsx and injected into SpeechService.ts.
• Audio is chunked (5s), converted to WAV snapshots, and transcribed via Whisper context.
• Silence tokens are stripped and transcript text is forwarded into existing analysis engines.
• TTS model loading is intentionally out of the critical path for now (placeholder/TODO remains).

---

Architecture Overview

High-level flow:

1. SpeechService.ts
   • Live audio capture
   • 5-second chunking
   • WAV snapshot temp file creation
   • Whisper transcription with domain prompt bias
2. SessionEngine.ts
   • Receives transcript chunks
   • Runs classifier/analyzer pipeline
   • Produces tactical detections and suggestions
3. ModelService.tsx
   • Downloads model files
   • Initializes Whisper/Llama contexts
   • Tracks model readiness/progress
4. Local storage services
   • Persist settings, sessions, and replay data

All existing analytical logic in src/ai/* remains algorithmic and local.

---

Key Project Files

• src/services/SpeechService.ts
  On-device STT bridge using whisper.rn, chunking, transcript emission.
• src/services/ModelService.tsx
  Model download/init lifecycle with whisper.rn + llama.rn.
• src/services/SessionEngine.ts
  Live orchestration and analysis pipeline.
• src/ai/*
  Pattern/intent/scoring/replay engines.

---

Privacy Model

• Inference runs on-device
• Audio is processed locally
• Transcripts/sessions are stored locally
• No mandatory cloud API for core session flow

---

Requirements

• Node.js 18+
• React Native 0.83.1 toolchain
• Android Studio / Xcode setup per standard RN environment
• Physical device recommended for realistic audio + model performance

---

Setup

npm install

iOS

cd ios
pod install
cd ..

Then run:

npx react-native run-ios

Android

npx react-native run-android

---

Development

Start Metro:

npm start -- --reset-cache

---

Migration Summary (completed)

• Removed previous SDK package dependency lock-in
• Replaced STT runtime with whisper.rn
• Replaced LLM runtime with llama.rn
• Updated model lifecycle to direct file download/init flow
• Preserved app logic/UI/session behavior while modernizing runtime layer

---

Troubleshooting

iOS pods not installed

Run:

cd ios && pod install

Model not ready in live session

• Ensure model download finished in app flow
• Confirm storage permissions/device storage availability
• Retry model load from session entry path

Slow transcription

• Use a physical device
• Ensure no heavy background workloads
• Keep device thermals under control (charging + heat can throttle CPU)

---

Roadmap

• Reintroduce local TTS runtime backend
• Add richer model management (variants, switching)
• Improve multilingual STT support
• Expand LLM-powered tactical generation pathways

---

License

Internal/proprietary project context unless specified otherwise.