App — AI pipeline

The companion's AI runs on the iPhone through Apple MLX. No conversation content is sent to the cloud. Each turn flows through a fixed pipeline:

ContextBuilder → PromptPipeline (7 layers) → CompanionOrchestrator
   → MLX Gemma 4 (4-bit) → StreamSanitizer → ToolParser → ToolValidator

Model

A PLE-safe 4-bit Gemma 4 model handles text, audio waveforms and images in one runtime. Audio reaches the model as raw bytes — there is no transcription step, ever.

Performance

• Two-phase warmup pre-compiles Metal kernels and the full attention graph, so there's no cold-start lag on the first message.
• A session-level KV prompt cache diffs token IDs and re-prefills only the delta, dropping time-to-first-token from ~174 ms to ~30–50 ms on later turns.
• Streaming output is sanitized at three gates so raw tool markup never reaches the screen.

Memory

Long-term facts are stored on-device with two-pass deduplication — lexical (Jaccard) then semantic (cosine over 384-dim MiniLM embeddings). Retrieval is decay-weighted by confidence, recency, use and importance tier.