Electric Safe

Electric Safe is a fully on-device Android application that guides industrial electricians through the Lockout/Tagout (LOTO) safety procedure using a Vision-Language Model (VLM) running on the Qualcomm Hexagon NPU via ExecuTorch and QNN. How it works — 5-screen state machine: Start — Dark industrial UI with a single "Start Session" button and a model selector (SmolVLM 500M / InternVL3 1B). Add Documents — Import equipment manuals via the SAF PDF picker or use the bundled PowerFlex 753 VFD manual. PDFs are processed on-device using PDFBox with regex-based fault code extraction. Live AI Session — Push-to-talk speech input (offline, EXTRA_PREFER_OFFLINE) and on-device TTS. Upload a photo of the VFD fault display; the VLM reads the fault code (e.g. F071 OC1) and matches it to the manual's meaning. A live NPU: X ms latency readout shows real inference time. Guided LOTO — Four camera-gated verification steps: Breaker B-201 OFF, Breaker B-205 OFF, Lock & Tag applied, MCC Cabinet open. Each step requires the VLM to verify the correct breaker identity and state (OFF/ON) with a minimum 0.70 confidence threshold. Identity mismatches trigger a red "WORK BLOCKED" card with expected vs detected values. Permit — Generates a timestamped LOTO evidence PDF on-device using android.graphics.pdf.PdfDocument, embeds captured photos, and allows sharing via FileProvider. Architecture: Single-activity MVVM with state-machine navigation (no NavController). The CaptureDetectionSource interface abstracts detection — mock mode (clickable demo) and real VLM inference drive the exact same code path. Model files are side-loaded via adb push to external storage (industrial MDM pattern), falling back to bundled assets. Privacy: No INTERNET permission declared. Camera capture, VLM inference, PDF generation, speech recognition, and TTS all run locally on the Snapdragon Galaxy S25 Ultra. Nothing leaves the device.