Capture Narcos · self-hosted voice engine
A store-visit report filled out entirely by voice. Speech-to-text runs locally on this machine; prompts use the device's own speech. Zero vendor API, zero tokens, zero per-minute fees.
Tap to wake Jarvis
Try: “What is this system?” · “Start the store report”
After the report: tap a field to fix it, or say “change the email” · “send it” · “show my last report”
Type an answer, pick a field + models, and see how each parses it. (Voice still works in the normal view.)
Engine: faster-whisper small.en (int8, local) for speech · a local LLM (Ollama qwen2.5:14b, local, on the GB10 GPU) understands spelling cues, cleans names & emails, and maps free-form answers; simple fields parse deterministically. All on-box — nothing leaves the Spark.