Real-time video inference system for geriatric patient monitoring

"Orión Ve, No Interpreta" (Orion Sees, Doesn't Interpret)

Orion is an edge-first AI sensor that captures video streams and produces structured inference outputs for the Care Scene ecosystem.

Orion is NOT:

• ❌ A competitor to Frigate NVR (end-user product)
• ❌ A competitor to DeepStream/DL Streamer (monolithic frameworks)
• ❌ An interpretation or decision engine

Orion IS:

• ✅ A configurable "smart sensor" for distributed architectures
• ✅ Best-in-class for event-driven AI sensor deployments
• ✅ A building block for larger monitoring systems (Care Scene)
• ✅ Hardware-agnostic (ONNX enables GPU acceleration)

• Go 1.x: Orchestration, streaming pipeline, control plane
• Python 3.x: ONNX inference workers (YOLO11 person detection)
• GStreamer: Video capture from RTSP streams
• MQTT: Event-driven control and data distribution
• ONNX Runtime: Multi-model ML inference
• MsgPack: High-performance binary serialization (5x faster than JSON+base64)

1. Complexity by Design, Not by Accident - Attack complexity through architecture
2. Pragmatic Performance - Real-time responsiveness > completeness
3. Non-blocking Channels - Drop frames to maintain <2s latency
4. Hybrid Go-Python - Go for orchestration, Python for ML
5. MQTT-centric Control - Hot-reload without service restart
6. KISS Auto-Recovery - One restart attempt only

• Go 1.21+
• Python 3.10+
• GStreamer 1.20+
• MQTT broker (Mosquitto)

# Build binary
make build

# Run with default config
make run

# Run with debug logging
./bin/oriond --debug

# Run with custom config
./bin/oriond --config path/to/orion.yaml

Primary config: config/orion.yaml

instance_id: orion-hab-302
room_id: hab_302

camera:
  rtsp_url: rtsp://camera-ip/stream

stream:
  resolution: 720p
  fps: 30

models:
  person_detector:
    model_path: models/yolo11n.onnx
    max_inference_rate_hz: 1.0

mqtt:
  broker: tcp://localhost:1883

# Get status
mosquitto_pub -t care/control/orion-hab-302 -m '{"command":"get_status"}'

# Pause/Resume inference
mosquitto_pub -t care/control/orion-hab-302 -m '{"command":"pause"}'
mosquitto_pub -t care/control/orion-hab-302 -m '{"command":"resume"}'

# Change inference rate
mosquitto_pub -t care/control/orion-hab-302 -m '{"command":"set_inference_rate","rate_hz":2.0}'

# Hot-reload model size
mosquitto_pub -t care/control/orion-hab-302 -m '{"command":"set_model_size","size":"m"}'

See docs/DESIGN/ORION_2.0_PLAN_EVOLUTIVO.md for detailed roadmap.

• ✅ Fase 1 (v1.0 → v1.5): Foundation - Bounded contexts, single-stream, hot-reload
• 🔄 Fase 2 (v1.5 → v2.0): Scale - Multi-stream (4-8 rooms), resource management
• 📅 Fase 3 (v2.0 → v3.0): Intelligence - Cell orchestration, motion pooling

• C4 Model - Complete architectural views
• Big Picture - Orion 1.0 overview
• Evolution Plan - 3-phase roadmap

• CLAUDE.md - AI-assisted development guide
• Design Manifesto - Design philosophy

"Las buenas prácticas son vocabulario de diseño - las practicas para tenerlas disponibles cuando improvises, no porque la partitura lo diga."

• ✅ DDD for bounded contexts clarity
• ✅ SOLID where it matters
• ✅ Pragmatism for utilities
• ❌ NO Hexagonal "because best practice"
• ❌ NO DI everywhere "because SOLID"

Read the scales, improvise with context.

This is a B2B consultative product. Contact Visiona for collaboration.

1. Read CLAUDE.md + Big Picture.md before coding
2. Attack complexity through architecture, not code tricks
3. Validate compilation as primary test
4. Document architectural decisions (ADR style)

• Co-authored by: Gaby de Visiona <noreply@visiona.app>
• Focus on "why" rather than "what"

Proprietary - Visiona © 2024-2025

Part of the Care Scene ecosystem:

• Orion - Smart sensor (this repo)
• Scene Experts Mesh - Event interpretation
• Room Orchestrator - Resource coordination
• Temporal Supervisor - Continuous learning

Built with pragmatism, designed for scale. 🚀