What's actually happening in your tank
Your reef never sleeps.
Neither does ReeferVision.
Computer vision + continuous telemetry across your whole water system — RO production, mixing, quarantine, top-off, and the display itself — fused into one model of your reef. Catch problems hours before you'd notice at the glass.
Why this exists
Reef tanks don't crash suddenly.
They crash while you're at work.
of new reef systems experience a catastrophic crash within the first 18 months, most traced to undetected parameter drift or equipment failure.
average livestock loss in a single crash event for serious hobbyist systems. Commercial losses run 10–50× that.
typical window between the first measurable anomaly and visible coral stress. By the time you see it, you're late.
You're testing at 11pm with a Hanna checker and a headlamp. Your alkalinity was drifting for two days before that test. The data was there. Nothing was reading it.
What ReeferVision watches
One model of your entire reef.
Every reading feeds a per-tank baseline — so when something's wrong, the alert is about your tank, not a generic threshold someone else set.
AI Coral Vision
A fine-tuned vision model scores each colony for polyp extension, tissue color shift, and recession — every 15 minutes, against your tank's own photographic baseline. It distinguishes normal nighttime retraction from genuine stress response.
Vertex AI vision · training set grows with every beta tank
Predictive Alerts
Time-series models learn your dosing rhythm, evaporation curve, and consumption rates, then flag drift before it crosses a danger line. Detects alkalinity sliding 0.3 dKH off-trend before it kills your SPS — not after.
BigQuery ML · forecasting on 10-second telemetry windows
24/7 IoT Sensor Fusion
pH, temperature, salinity, alkalinity, nitrate, and phosphate streamed continuously and cross-validated against each other. When a probe fails or drifts out of calibration, ReeferVision tells you it's the probe — not your tank.
Pub/Sub event pipeline · sub-minute ingestion latency
Fleet Intelligence
Anonymised anomaly patterns across the network quietly sharpen every tank's model. When a bad salt batch or a failing heater model starts showing up in other systems, your alerts improve before it reaches you.
Opt-in, anonymised · cross-fleet anomaly signatures
Digital Twin
A 3D flow-and-light map of your tank, built from your wavemaker positions, fixture, and camera view. Move a pump in the app and watch the flow map react — before you move it in the water.
interactive scape builder · PAR & flow estimates per spot
Guided Workflows
Acclimation timed by actual bag readings, water changes from a matched mix vat, placement recommendations per species. The hands-on moments where tanks go wrong — walked through, step by step.
live acclimation timer · per-species placement advice
Setup
Online in an afternoon.
Useful from day one, calibrated by week two.
Connect sensors + camera
The ReeferVision hub pairs with your existing probes (Apex, Hydros, GHL supported) plus our calibrated camera module. No proprietary lock-in on hardware you already own.
setup time ≈ 40 min · WiFi or Ethernet
AI learns your baseline
Over 14 days the model maps your tank's daily pH swing, photoperiod response, dosing signature, and each colony's normal appearance. Your reef becomes its own reference.
per-tank model · no generic thresholds
Get alerts before escalation
Drift, stress, or equipment anomalies trigger graded alerts — watch, warn, act — with the evidence right there: the trend chart, the camera frame, what probably caused it.
push · SMS · email · webhook
How it all works
From probe to phone,
and back again.
No magic. Sensors and a camera at the tank, a managed cloud pipeline, models that learn your system, and a loop back down to the hardware when something needs doing.
Local sensors & hub
Probes, level and flow sensors, leak sensors, and the camera all feed the ReeferVision hub at the tank. The hub buffers everything locally and runs the safety-critical rules itself — stop the return pump, lock out the ATO — so flood protection keeps working even when your internet doesn't.
edge device · local rule engine · works offline
Up to the cloud
Readings and camera frames travel encrypted to a managed ingestion service, land on an event backbone, and settle into a time-series warehouse — frames into object storage. Nothing for you to maintain, and it costs nearly nothing while a tank is quiet.
Cloud Run ingestion · Pub/Sub · BigQuery · Cloud Storage
The AI layer
Two kinds of models. A vision model scores each coral colony against your tank's own photographic history. Time-series models learn your daily pH swing, dosing rhythm, and evaporation curve, then forecast forward and flag anything off-trend — including equipment failure signatures like a heater starting to oscillate.
Vertex AI vision · BigQuery ML forecasting · per-tank baselines
Useful information, not noise
Everything the models find gets turned into things you can act on: graded alerts (watch / warn / act) with the chart and camera frame attached, maintenance clocks, consumable forecasts, and a dashboard that syncs to your phone in real time.
Firebase sync · push notifications · evidence attached to every alert
Automation closes the loop
For flood-risk and livestock-risk events, decisions flow back down to the hub: throttle the return pump, pause dosing, lock the top-off. The same safety rules are mirrored locally on the hub, so the protective actions never depend on a working connection.
cloud decisions → hub commands · safety mirrored at the edge
Live demo · sample data
It's not just the display tank.
It's the whole water system.
A reef is a chain of processes — RO production, mixing, quarantine, top-off, the display itself. A crash can start anywhere in that chain. ReeferVision watches all of it as one system. Here's what a real setup looks like in the dashboard:
Where every drop starts. ReeferVision tracks TDS in and out, membrane pressure, and production rate — and learns when your DI resin or membrane is on the way out.
model note: membrane rejection normal · DI resin at ~78% capacity, replacement forecast 6–7 weeks · filtered tank covers ~5 days of top-off · waste tank empty reminder at 85%
Water change water, monitored while it mixes and brought to match the display before it ever touches your tank. No more guessing whether the batch is ready.
model note: batch matched to display, safe to transfer · salt consumption learned from your water-change rhythm · reorder reminder lands ~5 days before you'd run out
QT systems are unstable by design — small volume, no mature filtration, medicated water. The vision model scores corals in treatment, and the float valve gets cleanliness tracking, because a gunked-up QT float is how hospital tanks overflow.
watch item: trace ammonia rising since yesterday's feeding · suggested: small water change from matched mix vat (STN-02) · coral in treatment trending up from 79 over 8 days · float clean scheduled with next water change
The quiet failure point in most systems. ReeferVision tracks reservoir level, fill events, and pump behaviour — a stuck float valve looks very different from normal evaporation, and the model knows it.
model note: fill cadence matches learned evaporation curve · reservoir lasts ~8 more days · refill reminder scheduled
The engine room. Sock loading, skimmer behaviour, and return pump draw all have learned signatures — and every serviceable part in the system gets a maintenance clock, so cleans and seal replacements happen on schedule instead of after a failure.
model note: clocks aren't fixed timers — sock loading is read from the level differential across it, and the schedule adjusts to how your system actually fouls
And when something does go wrong, it reaches your phone — graded, with the evidence.
Overflow restrictions, leaks, stuck floats, drain blockages — the failures that flood floors don't announce themselves. ReeferVision grades every alert so you know at a glance whether it can wait until you're home:
- WatchSomething's drifting. Logged and tracked — no action needed yet, but the model is paying closer attention.
- WarnIntervene soon. A trend is heading toward a threshold. You get the chart, the camera frame, and the likely cause.
- ActRight now. Flood-risk and livestock-risk events. Where it's safe to, ReeferVision responds first — throttling the return pump or locking out the ATO — then tells you what it did.
This is sample data, cycling live so you can see how the dashboard behaves — the numbers are realistic but not from a real tank (yet). Every station feeds the same per-system model, so an anomaly upstream (rising TDS, an off-spec mix batch) is connected to what it will mean downstream in the display.
Guided workflows
Monitoring is half of it.
The other half is doing things right.
The moments where tanks actually go wrong are the hands-on ones — adding livestock, deciding where a coral goes. So the app walks you through them, using what it already knows about your water and your layout.
Tell the app what you're adding and it builds the acclimation plan from live readings — how long the bag floats is decided by the actual temperature difference, not a guess, and the drip rate is set from the salinity gap between bag and tank.
- ✓ Lights dimmed to 20% · flow pumps paused near arrival zone done
-
2
Float bag — equalising temperature
finishes when bag ΔT < 0.2°C, not on a fixed timer 11:42 - 3 Drip mix — slow 1:1 water blend over ~40 min queued
- 4 Placement — recommended spot ready (see tank map) queued
app note: phone buzzes when each step is actually ready · no more standing over the bag with a kitchen timer · full log saved to the livestock record
The app builds a flow-and-light map of your tank from your wavemaker positions, light fixture, and camera view — then recommends where a new coral should go based on what that species actually wants.
app note: the map updates when you move a pump or change the light schedule · PAR figures are model estimates — calibrate with a one-off PAR meter session for precision
Limited beta · Cohort 3 opening soon
We're testing with a small group first.
ReeferVision is early. We're not selling subscriptions yet — we're putting hardware in tanks, training models, and learning what actually matters. When we do launch paid plans, beta members will get founding pricing locked in and a real say in what the tiers look like. For now, it's just the waitlist.