The brain · two-stage AI
One fast model for every clip. One careful model for the hard ones.
Every motion event runs through Stage 1: a fast, cheap vision model that clears deliveries, posties, residents and possums in a fraction of a second. Only the clips it isn't confident about escalate to Stage 2, a slower reasoning model. That gate is why the pipeline scored 92% on real CCTV without paying Sonnet's price on every postie.
Claude Haiku 4.5
Anthropic's fast vision model. Reads six frames from every 30-second clip and returns a one-paragraph plain-English summary plus a tier verdict with a confidence score. Latency under three seconds end-to-end.
- Runs on every motion event, no exceptions
- ~$0.006 per clip, including the six-frame vision payload
- Model id:
claude-haiku-4-5 - Prompt-cached system block cuts repeat cost ~70%
Claude Sonnet 4.6
Anthropic's reasoning-grade model. Fires only when Stage 1's confidence drops below 0.90. It gets the same six frames plus Stage 1's verdict and reasoning, and it can disagree. The vehicle-ramming-gate case in the scorecard is the canonical recovery.
- Fires on ~12% of clips in the May 2026 validation run
- +$0.013 on top of the Stage 1 cost
- Model id:
claude-sonnet-4-6 - Blended per-clip cost across the pipeline: $0.019
The 0.90 gate is deliberately conservative. Better to pay Sonnet a fraction of a cent more on an ambiguous clip than miss the one frame that mattered. As the corpus grows past fifty frames the gate may move, and every change gets published here with the date it changed.
The edge · hardware at your home
A small box on your shelf does the early thinking.
Sentry doesn't stream your cameras to the cloud. A Vigil Guard bridge agent lives on a Raspberry Pi at your house, behind your network. It captures only the clips your camera's own motion sensor already flagged, applies your mask zones, and decides what leaves the building. The cloud sees a flagged clip, never a live stream.
Raspberry Pi
Runs the Vigil Guard bridge agent on your network. A Pi 5 by default, or a Pi 4 where you already run a Unifi controller. It subscribes to your camera's existing motion events, captures the 30-second clip, applies your mask zones, and decides whether the cloud needs to see it at all. The Pi never streams unmonitored frames anywhere.
Coral USB Edge TPU
An optional AI accelerator on the roadmap, plugging into the Pi via USB-C. Where fitted, it runs a tiny on-device vision pre-filter that drops obvious non-events (leaves, rain, the cat on the porch) before they reach the cloud. The end-to-end pipeline runs without it today; the cloud verdict comes from Claude.
Your existing camera
Today: Unifi Protect (G3/G4/G5), validated end-to-end. On the roadmap: other RTSP cameras (Reolink, Hikvision, Dahua, Amcrest) as each is validated. Your camera doesn't need to know Sentry exists; the bridge agent reads its motion stream like any other client.
Why we don't sell our own camera. Camera hardware is a commodity. The work that matters is the layer that reads the footage, and that layer is the same whichever IP camera you bought last year. Locking members into a Vigil Guard camera would mean another box on your shelf, another firmware to maintain, and a slower upgrade path the next time a better RTSP camera ships. The four-pillar membership pays for the thinking layer; the camera you already own stays yours.
Storage · Australian-only
Clips never leave the country. Frames never get stored at all.
Sentry stores the source clip, never the individual frames. After Claude returns its summary, the six frames the model read are discarded. Only the clip itself persists, in Australian storage, on your retention setting. This mirrors the privacy section on /sentry, repeated here for engineering completeness.
- Object storage A Railway persistent volume in Sydney. Clips serve only through signed URLs (HMAC, with an expiry and signature on every link), private and scoped to your account. Clip-only persistence: the six frames the model read are never written to disk.
- Default retention 30 days. Member-controllable between 7 and 90 days from the member app. After retention expires, clips are deleted from storage and from backups.
- Compute region Railway services in Sydney. The vision worker, ingest gateway and escalation router never run outside Australia. Anthropic API calls leave Australia only for the millisecond round-trip; under Anthropic's enterprise terms, customer data is not used for training.
- Mask zones Polygon regions you draw before Sentry analyses your first clip. Applied at the vision-worker layer before any Anthropic API call. Masked pixels never reach the model.
- Deletion on request Email privacy@vigilguard.com.au. Full erasure of every clip, every metadata record and every audit-log entry within 72 hours, with a written audit log of the deletion as proof.
- Append-only audit log Every clip Sentry analysed, every tier it assigned, and every alert it raised to you and your contacts. You can request a full copy of your audit log at any time.
Architecture · every hop
The path from your camera to your phone.
Every box in this diagram exists in code today. The blue path is the common case: Stage 1 is confident, and the clip files quietly to your app feed. The crimson path is the careful one: Stage 1 is uncertain, Stage 2 takes a second look, and within seconds you, your four contacts and your street get alerted with your live location.
No part of this diagram is theoretical. Every box maps to a service or piece of hardware that exists today, validated end-to-end on the May 2026 corpus. The escalation router that sorts Tier I / II / III is the same code the founding-cohort closed beta runs against. No clip auto-dials anyone: a Tier III raises a member alert, and you decide whether to hit panic or call 000.
The phone · native client TestFlight now · App Store submission pending
The last hop lives on your phone.
Every alert ends at a native iOS app, built in Expo and React Native. It's on TestFlight with founding-cohort testers today, with App Store submission pending. The software runs in every Australian postcode; patrol lights up state by state where we're licensed.
Home is watched
A live status panel that tells you, in plain language, whether your watch is running right now and what it has cleared tonight. No dashboard to read, just the one answer you want.
Protected-night streak
A running count of the nights your home has been watched without an event that needed you. It builds the same way a habit does, and gives you a reason to keep the watch on.
Neighbourhood radar
A live suburb map of nearby activity: your reports, your neighbours' flags, and reported crime placed by real distance from your postcode. Built with native maps, no third-party tracker.
Night-watch mode
An ambient mode for the hours that matter. The app settles into a calm dark view while the watch runs, so a glance tells you everything is quiet without waking the whole screen.
Beyond Sentry · the live pillars
Sentry is one pillar. Three more run today.
Sentry watches your cameras. The rest of the stack watches your people. Panic, Lookout and Guardian Walk run as national software in every Australian postcode, with the same honesty rules as the rest of this page.
Hold three seconds
Hold the panic switch for three seconds and the app alerts your four nominated contacts with your live GPS. On the Guardian tier it also fans out to your street and neighbour network with your location. The free tier alerts your contacts only.
- Parallel call tree to four contacts · live location attached
- Guardian tier adds the street fan-out; free tier is contacts only
- Vigil Guard never dials 000. A one-tap call-000 button is yours to press in a life-threatening emergency
One tap, behaviour only
Flag suspicious behaviour in one tap, never appearance, race or clothing. Three confirmed flags in an area trigger a cluster alert. A neighbour-verification layer lets members vouch for real residents, building a verified constellation that makes the alert network harder to spoof.
- Behaviour-only flags · appearance fields don't exist by design
- Three confirmed flags in an area raise a cluster alert
- Vouched-neighbour constellation: more verified residents, less spoofable network
A watched walk home
Set an ETA and the app pings your live location as you walk. Miss your check-in or go silent and your contacts and street get your last known location. Steady runs alongside it as a staged companion that escalates over a walk and can fire the panic alert.
- Live-location pings against an ETA · missed check-in fans out your last location
- Steady escalates in stages: walk with me · stay close · get ready
- Free shows your dot; the watched walk and auto-escalation are the Guardian layer
Crime Watch · the data
Where the incident feed comes from.
Crime Watch is built to run on real Crime Statistics Agency Victoria data, broken out per council area and postcode, surfaced on /watch and /safety. To keep the "no black box" promise honest: the live public ticker still draws on attributed news incidents while we move it onto a fully licensed crime feed. We will name the source the day that swap lands.
Open questions, honestly
The three things people ask first.
Engineering questions get engineering answers. If yours isn't here, email hello@vigilguard.com.au and a founder answers.
Why Anthropic, and not OpenAI or Google?
Three reasons. First, Anthropic's enterprise terms forbid training on customer data: there is no path under which member footage trains a future model. Second, the Claude model family's calibration on refusing to confidently mis-classify ambiguous frames is, as of the May 2026 validation run, the best fit we measured for "default toward maybe / Tier II when uncertain, never default to a confident Tier III false alarm." Third, the two-stage Haiku → Sonnet design relies on both models being from the same vendor's lineage so prompts and tool schemas remain consistent. None of these are forever-decisions; if a competing vendor's terms and calibration outperform Anthropic on a future audit, we will publish that audit and migrate.
Why a Pi at my house instead of running everything in the cloud?
Three reasons again. First, latency: the on-device pre-filter cuts the time between motion and Stage 1 verdict by ~700ms on average. Second, cost. Most motion events (rain, leaves, your cat) are dropped before they leave your network, which keeps cloud compute low enough to run Sentry sustainably for the founding cohort. Third, and most important, you keep the frames on your network. The bridge agent only uploads clips your camera's motion sensor flagged and the on-device pre-filter judged worth a second look. Frames that never leave your house can never leak from our cloud.
Is the source open?
Not at the moment. We may open-source the bridge agent later, since it runs on your hardware and you have a reasonable interest in reading what it does. The cloud side, the escalation router and the prompt blocks, stays closed for now, like any alerting backend. What is public: the validation corpus and its verbatim per-frame verdicts (see the scorecard), the model identifiers, the cost figures, this architecture diagram, and every change to the confidence gate over time, dated. Sentry never dials 000 or any emergency service on your behalf. A Tier III raises an alert to you, your four contacts and your street; the call to emergency services stays yours.
Sentry · founding-cohort closed beta
Now you know what's running. Want it on your home?
Own a compatible camera already? Put your email in. The first fifty closed-beta invites go to founding-cohort members in the three pilot postcodes, then we expand. Full membership detail for Watch, Patrol and Estate sits on the pricing page.
No spam. We tell you when there's a slot: that's it.
Thank you. We'll be in touch when a slot opens.
Or read more · Sentry overview · Validation scorecard