FishTank

From automation
to process autonomy.

FishTank is the window on our Agentic AI: a modular software agent that does not just execute a program, but reads the plant, decides and acts on its own, and tells people what it did. We proved it on a 250-litre planted tank in our Flero lab, running day and night on NVIDIA Jetson. The aquarium is the test bench; the agent is for any plant where you want to move from automation to autonomy.

Coming soon on the App Store · iPhone and iPad
Watch the tank live Bring it to your plant →
The FishTank aquarium in the Railab laboratory: discus, cardinal tetras and plants
The tank, from the camera the agent watches it with.
What it does

What the agent does on its own, here and on your plant.

LIVE
Every value, live
Temperature, pH, conductivity and organic load with their targets, the state of every device and the last 24 hours, open to anyone.
SCHEDULERS
Day and night
Heater, CO₂, filter and lights follow their own schedulers; a value changed from the app stays after a restart.
VISION
It sees the fish
A neural network running on the Jetson recognises the fish in the camera image and counts them, frame by frame.
MAINTENANCE
Nothing forgotten
Water tests, water change and sensor care come back when they are due; whoever does them marks them done from the app.
STAFF
Who did what
Railab staff log in with personal credentials and Face ID: every command and every change is recorded under their name.
PRIVACY
Open, not watched
Anyone can follow the plant without an account. The app collects no data from those who use it without logging in.

Terms of use · Privacy policy

How it reasons

Supervision, not command.

Agent loop: exteroceptive and proprioceptive sensors, perception with a model of itself and of the world, inference, decision and action on the world and on itself, new state, continuous cycle
The agent in its environment: it perceives, reasons, acts on the world and on itself, and checks the new state.
PERCEPTION
Model of the world, model of itself
Sensors and a camera with a neural network describe the tank; the agent also tracks the state of its own devices.
SUPERVISION
Targets, not valves
The agent sets setpoints and hysteresis; low-level control and protections stay with each device.
COUPLINGS
It sees the whole
CO₂ solubility depends on temperature and plants use CO₂ only under light: pH, temperature and photoperiod are coordinated.
COMPOSITE CONDITIONS
What a single controller misses
pH falling fast, CO₂ valve open and lights on: overdose risk, critical. Climate on but water off target for 20 minutes: device fault.
CLOSED LOOP
A command sent is not a command done
Every action is checked against its effect. Context counts: after feeding, the conductivity peak is expected and raises no false alarm.
PERIMETER
Responsible autonomy
Only reversible actions; graded severities with escalation; peer agents with no master; gaps in the history stay visible. Everything else goes through a person.

No language model in the control loop: agentic does not mean LLM. The same architecture runs a robotic cell; only sensors and actuators change.