Built for the frontline.
Powered by Edge AI.

Two native iPadOS applications purpose-built for fire, EMS, and investigation. Offline-first, LiDAR-enabled, and driven by on-device intelligence.

Product OmniApp / OmniHealth

Clinical Documentation & Fire Investigation Suite

Zero-latency field operations. On-device AI. Full NEMSIS 3.5 / NERIS compliance.

Built for iPad Pro M‑series silicon. OmniApp combines Qwen 2.5 LLM reasoning, MedGemma vision analysis, and WhisperKit voice transcription into a single offline‑first overlay. From AR scene capture to cryptographic chain of custody — everything runs locally, even in dead zones.

🧠

Edge AI Orchestrator

Qwen 2.5 (LLM) + MedGemma (VLM) + WhisperKit running concurrently on Apple Unified Memory via MLX / Metal. KV cache serialization for near‑zero TTFT.

📐

AR Scene Scanner

RoomPlan API generates 2D floor plans and 3D structural meshes. Hybrid Sketchpad overlays Apple Pencil PKCanvasView for burn‑pattern mapping.

📦

3D Evidence Capture

ObjectCaptureSession photogrammetry produces interactive .usdz models. NFC / QR chain‑of‑custody ties evidence cryptographically to AR anchors.

🏗️

AR Pre‑Planning

Tap‑to‑tag NFPA/ICS symbology (FDC, Knox Box, FACP) onto RoomPlan meshes. ICS Command Handoff exports spatial anchors to the command board.

📡

Bluetooth 6.0 & P25 Bridge

Channel Sounding for cm‑accurate crew ranging. Graph SLAM delivers directional MAYDAY vectoring. Direct integration with Motorola APX radios.

🗺️

ArcGIS & Spatial Bridge

Embedded Esri maps with offline .geodatabase layers. Georeferences RoomPlan coordinates to global GPS, publishing scans as spatial feature overlays.

📋

NEMSIS 3.5 / NERIS Auto‑Scribe

Qwen 2.5 ingests WhisperKit transcripts, MedGemma visual evaluations, and weather back‑cast data to draft compliant "Origin & Cause" narratives. Validates against state datasets (e.g., Maryland MIEMSS).

Product OmniPD Tactical Incident Reporting

Command & Reporting for First Responders

Offline‑first, Swift‑native, optimized for M‑series iPad Pro.

OmniPD delivers CAD summarization, voice‑driven incident scribing, and LiDAR‑based pre‑planning in a streamlined bento‑grid interface. Built for the officer in the command vehicle.

🎙️

AI Speech & CAD Pipeline

Qwen 2.5 7B (4‑bit MLX) summarizes raw CAD notes into 2‑sentence briefings. WhisperKit v3 Turbo runs on the Neural Engine for low‑latency field dictation.

📄

VisionKit OCR

On‑device scanning of driver's licenses, IDs, and HazMat manifests. Clean strings passed directly to Qwen for parsing.

🏛️

LiDAR Pre‑Planner

RoomCaptureView captures 3D building spaces. Exports .usdz models and JSON paired with GPS / heading metadata.

🥽

RealityKit Dual‑Mode AR

1:50 scale tabletop dollhouse mode anchored to command vehicle surfaces, or 1:1 real‑world walkthrough mode.

🗺️

ArcGIS Integration

Dark vector basemaps with active CAD call markers. 3D building pre‑plan overlays placed directly onto GIS locations.

🔥

Thermal & Memory Optimizations

Adaptive throttling steps down ARKit polling and caps token output during .serious thermal states. GRDB with WAL + mmap reads HazMat tables directly from disk.

📱

Bento UI / UX

Asymmetric SwiftUI grid with 6×2 hero tile for live briefings. Compact ShotSpotter banner with swipe‑to‑dismiss. Streamlined Action Deck consolidating camera, interpreter, and evidence tools.

⚙️ Shared Architecture

Both applications share a common foundation of on‑device AI, spatial intelligence, and enterprise‑grade compliance. Everything runs locally — no cloud required.

🧠

MLX Edge AI

Qwen 2.5, MedGemma, WhisperKit — all on‑device via Metal / Neural Engine.

🔒

Offline‑First

Zero dependence on cellular or Wi‑Fi. Full functionality in dead zones.

📡

LiDAR + ARKit

RoomPlan, ObjectCapture, ARWorldMap — spatial intelligence baked in.

🌡️

Thermal Management

Adaptive throttling keeps displays bright in vehicle mounts and high‑heat environments.

📋

NEMSIS / NERIS

Fully validated schema output for state reporting (MIEMSS / eMEDS / ImageTrend).

Zero‑Copy Memory

IOSurface → MLXArray direct mapping. No CPU‑to‑GPU RAM duplication.