[!NOTE] Project Status: Active Prototyping & Edge Hardware Phase Gefährte (German for “Companion”) is currently in active hardware and software prototyping. Built German-first on an NVIDIA-powered edge device, Gefährte delivers empathetic voice companionship, camera-free mmWave radar fall detection, and conversational medical reminders under an absolute architectural guarantee: all data is processed locally and will not leave the Haus.
Product Vision: German-First Aging Well
Demographic change is accelerating across Europe, with Germany at the forefront. Millions of older adults live independently at home, navigating social isolation while families live across distant federal states and care institutions grapple with acute staffing shortages (Pflegenotstand).
Existing solutions repeatedly fail because they force an unacceptable compromise between safety and personal dignity:
- Invasive Optical Cameras: Seniors overwhelmingly reject video surveillance in their private quarters, bedrooms, and bathrooms.
- Wearable Panic Pendants: Fall detection buttons are routinely left on bedside tables, forgotten during night walks, or taken off before showers where slip-and-fall incidents most commonly occur.
- Generic Cloud Chatbots: English-centric cloud AI systems introduce network latency, misunderstand regional German idioms, and exfiltrate intimate conversations to offshore server farms.
Gefährte is conceived from a radically different philosophy: a dignified, ambient, voice-first companion built German-first for the home. By combining an NVIDIA-powered embedded platform (GPU + CPU), a custom localized AI model, 60GHz mmWave (micro-wave) radar fall detection, and proactive medical reminders, Gefährte offers round-the-clock safety and warm conversational presence—without cameras, without wearable tags, and with the ironclad guarantee that personal data never leaves the Haus.
Core Capabilities
1. German-First Natural Voice Interface & Custom AI Model
Loneliness cannot be solved through complex touchscreens or smartphone apps. Gefährte eliminates screens as a prerequisite for companionship.
- German-First Linguistic Tuning: Gefährte is built from the ground up for German speech patterns, honoring regional dialects (Bairisch, Schwäbisch, Plattdeutsch), formal/informal address (Sie vs. Du), and respectful conversational pacing suited to senior cognitive comfort.
- Custom On-Device AI Model: Rather than routing sensitive conversations through generic public LLM endpoints, Gefährte runs a custom-trained, fine-tuned localized language model directly on the hardware. It possesses deep eldercare domain awareness, gentle active listening abilities, and natural conversational cadence.
- Pure Voice Interface: Far-field microphone beamforming and acoustic echo cancellation allow older adults to speak naturally from across the room without holding a device or straining their voice.
2. NVIDIA-Powered Dual-Engine Compute (GPU + CPU)
To deliver instantaneous voice interaction and real-time radar point-cloud filtering without relying on external internet connectivity, Gefährte is powered by a dedicated NVIDIA edge compute platform:
- Dedicated NVIDIA GPU (Tensor Cores): Accelerates on-device custom neural model inference, sub-second neural Speech-to-Text (STT), naturalistic neural Text-to-Speech (TTS), and real-time deep learning classification of mmWave radar point clouds.
- High-Performance Multicore CPU: Orchestrates low-latency Go system daemons, manages real-time audio DSP, executes local scheduling engines for medical reminders, and handles hardware I/O with absolute determinism.
- Complete Edge Autonomy: System reasoning and sensor monitoring remain 100% operational even during total internet or Wi-Fi disruptions.
3. Absolute Privacy: “Data Will Not Leave the Haus”
Privacy in German eldercare is non-negotiable. Gefährte implements an uncompromising architectural boundary: Data will not leave the Haus.
- Zero Cloud Audio Streaming: Continuous room audio is processed strictly within volatile local RAM on the NVIDIA CPU for local wake-word detection. Raw speech audio is never streamed to external cloud servers.
- Local Inference & Vector Graph: Custom AI model weights, conversation histories, and episodic memory graphs reside entirely within local encrypted NVMe storage inside the device.
- German DSGVO / EU GDPR Compliance by Design: Because personal health observations, voice transcripts, and movement habits remain within the user’s physical four walls, Gefährte eliminates the risks of data breaches, third-party profiling, or unauthorized surveillance.
4. mmWave Radar Fall Detection with Zero Cameras
Falls are the single greatest threat to independent living for older adults, with the majority occurring in bathrooms and bedrooms where cameras are fundamentally unacceptable.
- 60GHz Micro-Wave (mmWave) Radar: Emits safe, high-frequency electromagnetic radar pulses that continuously map 3D spatial points, tracking human posture, gait velocity, micro-movements, and sudden drops in elevation.
- Zero Optical Cameras: There are no camera lenses, no image sensors, and no video recording hardware on Gefährte. The system perceives room dynamics purely through privacy-preserving radar reflection signatures.
- Through-Obstacle & All-Condition Reliability: Unlike optical sensors, mmWave radar functions flawlessly in total darkness, through dense bathroom steam, and even when a fallen person is partially obscured by bedding or furniture.
- Instant Autonomous Escalation: If an anomalous fall trajectory followed by immobility is detected, Gefährte immediately asks aloud in gentle German (“Herr Weber, sind Sie gestürzt? Brauchen Sie Hilfe?”). If no response or confirmation of distress is received within a calibrated window, the device initiates an automated alert to designated family members or emergency services.
5. Conversational Medical & Routine Reminders
Managing complex medication regimens is a daily challenge for many seniors. Traditional pill alarms use jarring beeps that cause anxiety or are casually dismissed.
- Contextual Voice Reminders: Gefährte delivers proactive, conversational reminders tailored to personal daily rhythms (“Guten Morgen, Frau Becker! Haben Sie nach dem Frühstück an Ihre Herztablette gedacht?”).
- Verbal Adherence Tracking: The custom AI model understands verbal confirmations (“Ja, habe ich gerade genommen” or “Nein, bringe ich gleich mit”), logging adherence locally and offering respectful follow-up check-ins if a critical dose was missed.
- Holistic Routine Care: In addition to prescription medications, Gefährte provides gentle prompts for hydration on hot summer days, scheduled doctor visits, and mobility exercises recommended by physiotherapists.
6. Local Episodic Memory & Personalized Life Narratives
Generic assistants treat every dialogue as a blank slate. Gefährte leverages an on-device PostgreSQL and pgvector semantic memory layer running locally on the NVIDIA platform.
- Cherished Memories: Securely indexes family relationships (“Enkelin Sophie studiert in München”), career milestones, favorite German folk melodies, and beloved gardening routines.
- Grounded Recall: Weaves past conversations into future interactions naturally (“Frau Weber, hat Sophie sich am Sonntag aus München gemeldet?”) without hallucinating false memories or leaking personal biographies outside the home.
Architectural Breakdown
┌────────────────────────────────────────────────────────────────────────┐
│ Living Space & Senior Resident │
│ 🎙️ Mic Array 🔊 Speaker 🔘 Tactile Button 📡 60GHz mmWave │
│ (🚫 ZERO CAMERAS / NO LENSES) │
└───────────────────────────────────┬────────────────────────────────────┘
│ Analog Audio + Raw Radar Signals
▼
┌────────────────────────────────────────────────────────────────────────┐
│ NVIDIA-Powered Edge Hardware Platform │
│ │
│ ┌─────────────────────────────────┐ ┌──────────────────────────────┐ │
│ │ Multicore CPU Engine │ │ Dedicated NVIDIA GPU │ │
│ │ (High-Velocity Go Daemon) │ │ (Tensor Cores) │ │
│ │ - Local Wake-Word & Audio DSP │ │ - Custom Local AI Model │ │
│ │ - mmWave Radar Signal Pipeline │ │ - Local Neural STT & TTS │ │
│ │ - Fall Detection State Machine │ │ - Radar Point Cloud AI │ │
│ │ - Medical Reminder Scheduler │ │ - On-Device Vector RAG │ │
│ │ - Local Encrypted Storage │ │ - Low-Latency Token Stream │ │
│ └─────────────────────────────────┘ └──────────────────────────────┘ │
│ │
│ 🔒 STRICT PRIVACY SHIELD: "DATA WILL NOT LEAVE THE HAUS" │
│ All voice, radar, and memory processing strictly local │
└───────────────────────────────────┬────────────────────────────────────┘
│ Optional Local Wi-Fi / Encrypted Alerts
▼
┌────────────────────────────────────────────────────────────────────────┐
│ Caregiver & Family Portal (Angular + TypeScript) │
│ - Non-Diagnostic Wellbeing Summaries & Activity Rhythms │
│ - Emergency Fall Detection Notifications & Urgent Call Routing │
│ - Medication Adherence Confirmations & Routine Log │
│ - Family Memory & Photo Injection Studio │
└────────────────────────────────────────────────────────────────────────┘High-Velocity Tech Stack
- NVIDIA Edge AI (GPU + CPU): The hardware heart of Gefährte. The dedicated NVIDIA GPU accelerates deep learning inference for the custom local LLM, acoustic models, and radar point-cloud classification, while the multicore CPU runs compiled Go daemons and system services with minimal power consumption.
- Custom Local AI Model: A specialized conversational model fine-tuned for German eldercare, providing natural turn-taking, empathetic companionship, and respectful cognitive alignment without cloud API dependencies.
- mmWave Radar Sensors: 60GHz micro-wave radar modules capturing spatial point clouds for continuous, unobtrusive room occupancy, micro-mobility, and camera-free fall detection.
- No-Camera Architecture: Zero optical components engineered directly into the hardware bill of materials, guaranteeing absolute physical privacy across all household living spaces.
- Voice Interface & Local DSP: Hardware beamforming microphone array, acoustic echo cancellation, and on-device streaming speech recognition enabling effortless, hands-free voice dialogue.
- Medical & Routine Reminders Engine: Local deterministic cron scheduler and conversational confirmation loop tracking medication schedules, doctor appointments, and hydration.
- Go (Golang): Core language for the on-device system daemon. Go’s native concurrency primitives (goroutines and channels) ensure low-latency audio capture, radar signal processing, and state machine transitions.
- PostgreSQL & pgvector (On-Device): Embedded relational and vector store housing user profiles, medical reminder schedules, and semantic embeddings of personal memories.
- Angular & TypeScript: Enterprise-grade responsive portal for authorized family members and professional caregivers to inspect high-level wellbeing summaries and receive critical fall alerts.
Regulatory, Privacy & Responsible AI Design
Developing an ambient companion for German seniors requires adherence to the world’s most rigorous data privacy and ethical standards:
- Data Will Not Leave the Haus: Gefährte establishes an absolute technical boundary. Raw microphone audio, radar waveforms, personal diaries, and medical schedules are never transmitted to corporate servers, cloud AI providers, or third-party advertisers.
- Camera-Free Dignity: By relying exclusively on mmWave radar for fall detection, Gefährte eliminates the psychological burden and privacy invasion of optical cameras in bedrooms and sanitary facilities.
- German-First Cultural Alignment: The system is designed to respect German cultural norms around privacy, independence, and social dignity, speaking fluent German without foreign conversational mannerisms.
- Separation from Medical Device Regulation (MDR): Gefährte functions as an ambient wellbeing, social companionship, and emergency alert assistant. By avoiding prescriptive clinical diagnosis, the system remains a supportive household companion without clinical paternalism.
- Zero Model Re-Training on Private Data: User conversations and life stories are never ingested into global training corpora. The custom model weights remain immutable on the device, enriched only via local retrieval.
Roadmap & Prototyping Stages
┌───────────────────────────────────────┐
│ Stage 1: German Voice & Custom AI MVP │ ✅ German conversational model & semantic memory
└───────────────────┬───────────────────┘
▼
┌───────────────────────────────────────┐
│ Stage 2: NVIDIA Edge & mmWave Agent │ 🔄 Active prototyping (GPU+CPU, radar & Go daemon)
└───────────────────┬───────────────────┘
▼
┌───────────────────────────────────────┐
│ Stage 3: German Senior Living Pilots │ 📅 Community & caregiver validation across Germany
└───────────────────────────────────────┘- Stage 1 (Completed): German-first conversational architecture, custom model fine-tuning for empathetic senior dialogue, and episodic memory extraction with pgvector.
- Stage 2 (In Progress): Integration of the NVIDIA-powered hardware platform (GPU + CPU), 60GHz mmWave radar sensor integration for camera-free fall detection, local Go device daemon, and conversational medical reminder engine.
- Stage 3 (Upcoming): Field pilot studies with senior care organizations, assisted living facilities, and independent older adults across Germany to validate long-term social engagement and fall-detection efficacy.