SafeHaven:
Offline Entry-Point Security with Metis, Frigate, and Home Assistant
Hi everyone,
I’m excited to share my final submission: SafeHaven.
SafeHaven is a privacy-first, offline home security system built to monitor the states that most NVRs still do not understand well: door position, latch/lock position, and garage open/closed state.
Most video systems today are good at detecting people, faces, packages, or generic motion. Those problems already have many solutions. The gap I wanted to address is different:
Was the door left open?
Is the latch unlocked?
Was the garage left open long enough to become a real security risk?
That is the problem SafeHaven is designed to solve.
Video:
Repo:
https://github.com/shashibhat/SafeHaven_v2/tree/main/safehaven_v2
What SafeHaven does
SafeHaven uses local cameras and Metis-powered inference to monitor critical entry points around the home and convert low-level detections into semantic security events.
Instead of only saying “motion happened,” it can determine:
- latch_unlocked
- door_open
- garage_left_open
- gate_ajar
It then publishes those results into Frigate so the user gets a familiar NVR interface with snapshots, clips, timeline events, and review workflows.
Why this matters
A large share of preventable home security incidents come from unlocked doors, partially secured gates, or garages left open rather than sophisticated forced entry.
In practice, families often do not need another person detector. They need a system that answers:
- Did we actually secure the house?
- Was the garage left open after unloading?
- Was the latch unlocked for too long?
- Can I review exactly what happened afterward?
SafeHaven is built around that operational need.
Architecture
SafeHaven is designed as a market-ready local product architecture, not just a model demo.
Core components:
- Axelera Metis for edge inference
- Frigate as the NVR, evidence timeline, and operator UI
- SafeHaven Core for semantic state machines and event logic
- Home Assistant support for downstream automation and action workflows
The architecture separates concerns cleanly:
- Frigate handles camera ingestion, UI, clips, and snapshots
- SafeHaven Core evaluates security state over time
- Metis-backed inference provides the visual understanding
- Home Assistant integration enables real-world action
Frigate + Home Assistant product direction
One of the strongest parts of this system is that it fits naturally into tools people already use.
Frigate
Frigate is used as the primary NVR and event review interface.
SafeHaven pushes semantic events into Frigate so they appear like native reviewable timeline items.
Home Assistant
SafeHaven is also designed to work naturally with Home Assistant, using Frigate as the evidence layer and Home Assistant as the automation layer.
That creates a path toward a real product experience such as:
- alert if garage is open for more than X minutes
- notify if front-door latch remains unlocked
- trigger a light/buzzer workflow
- auto-close the garage through Home Assistant
- trigger a smart lock workflow
- support “auto-heal” actions when safe to do so
This is important to me because a production-ready home security product should not stop at detection. It should support closed-loop response.
Technical implementation
SafeHaven v2 is built around:
- Metis-first state semantics
- Frigate Create Event API integration
- a Frigate detector plugin path for Metis-backed inference
- bounded queues, stale-frame dropping, and Prometheus metrics
- camera-specific ROI processing
- robust semantic state machines for open/closed/ajar/locked/unlocked logic
The repository includes:
- metis-detector inference sidecar
- safehaven-core semantic engine
- frigate-metis-plugin
- Docker-based local deployment
- docs and architecture notes
What makes this different
The key differentiator is not object detection by itself.
The key differentiator is turning video into state-aware security workflows.
There are many systems that can say:
- a person is present
- a package is present
- motion is present
There are far fewer systems that can reliably say:
- the latch is unlocked
- the garage has been left open for 7 minutes
- the gate is still ajar
- this is now important enough to escalate and preserve as evidence
That is the product category SafeHaven is targeting.
Demo highlights
In the final demo, I focused on two practical scenarios:
- Latch unlocked detection
- Garage left-open countdown and event creation
The workflow shows:
- visual inference on the camera feed
- semantic state evaluation
- Frigate event creation
- event review in Frigate UI with media evidence
I also included the broader product narrative: most real security failures are not exotic attacks, but normal lapses in securing entry points.
Future work
The next steps for SafeHaven are clear:
- improve garage and door-state model coverage
- expand multi-entry-point support
- deepen native Home Assistant workflows
- expose SafeHaven events as a more polished Frigate/Home Assistant card experience
- add auto-heal actions for supported locks and garage controllers
- continue improving Metis-native detector integration
Closing
SafeHaven started from a simple question many people ask at night:
“Did we close the garage?”
The goal was to build a system that can answer that question locally, privately, and with evidence.
Thanks to Axelera Metis, Frigate, and the surrounding local-first ecosystem, I was able to build a foundation for a practical home-security product that goes beyond generic detection and into real-world state awareness.
I’d love feedback from the community, especially around:
- state-based security workflows
- Frigate + Home Assistant product patterns
- Metis-native detector integration
- local-first automation and response design