Using AI to Manage Aquariums

For the past several months, I’ve been running a 19-tank freshwater aquarium system — roughly 170 gallons spread across everything from a 75-gallon community to 2.5-gallon neocaridina breeders — with an AI assistant as my primary operations partner. Not as a novelty. As a genuine engineering tool.

This isn’t about asking ChatGPT what fish to buy. It’s about building a persistent, data-driven management layer on top of a complex biological system. Here’s how it actually works.

The System Architecture

Every tank in my system is tracked in a JSON-based state management file stored in Box.com. The STATE file is a complete census — every animal, every tank, every piece of hardware, every pending change. It follows a commit-based versioning system (currently past commit 260) with audit logs, delta tracking, and cross-file validation.

Supporting files include a LIBRARY (species catalog with scientific names, bioload coefficients, and strata assignments), a MODELS file (mathematical formulas for bioload, filtration capacity, and evaporation), and subsidiary files for subsystems like automated top-off and feeding schedules.

The AI doesn’t just read these files. It fetches them from cloud storage, validates them, performs calculations against them, and pushes updated versions back. When I say “add 4 babaulti shrimp to the Chilly tank,” the AI pulls the current STATE, applies the delta, recalculates bioload, checks capacity headroom, and prepares a commit — all before I confirm the upload.

Bioload Modeling

This is where the engineering gets interesting. I developed a custom bioload unit system (BCU — Biological Capacity Units) calibrated against a Bristlenose Pleco as the 1.0 baseline. Every species in my catalog has a v2_unit value derived from metabolic scaling (mass^0.75), adjusted for feeding behavior and waste production.

The AI runs these calculations on demand. When I’m considering adding a school of Vietnamese Cardinal Minnows to a new tank, it pulls the species coefficient from the LIBRARY, multiplies by quantity, sums with existing livestock, and compares against the tank’s total filtration BCU to produce an RLI (Relative Load Index). Green means headroom. Yellow means slow down. Red means stop.

The model has been validated against published metabolic scaling research (King 1996) and the calculations at AquariumScience.org. The AI performed that validation — pulling external sources, comparing coefficient-by-coefficient, and flagging discrepancies.

Filtration Engineering

Every filter in the system has a cataloged BCU rating. When I started converting filters from stock media to staged Poret foam (graded-density foam that eliminates bypass flow), the AI calculated the BCU uplift for each filter type. An AquaClear 110 went from 17.0 BCU with stock media to 24.0 BCU with Poret staging — a 41% gain from a media swap alone.

When I considered adding a second 75-gallon tank, the AI ran side-by-side comparisons of different filter combinations — tandem canisters versus canister-plus-HOB, all normalized to Poret-staged BCU values — so I could see exactly how close a two-filter setup could get to my existing three-filter tank’s capacity.

Automated Top-Off Calculations

My system uses four Jebao 4-channel dosing pumps feeding from a single RO/DI reservoir to compensate for evaporation across all tanks. The AI calculated every dosing schedule from first principles — surface area, ambient humidity (30% RH winter, 50% summer), temperature, and cover percentage — producing per-tank, per-season dosing plans with safety derating for a 3-week manual top-off cycle.

It figured out the optimal dose splitting (multiple small doses over an hour rather than one large dose for pump accuracy and flood safety), calculated timing intervals to prevent channel overlap, and generated the complete programming table for all 16 channels.

Custom Fabrication Support

When I needed custom acrylic covers for rimless tanks, the AI acted as a mechanical engineering consultant. It calculated deflection limits for 1/8″ acrylic spans, designed multi-segment covers with bent reinforcing lips, specified bend radii to avoid crazing, and produced complete cut-and-bend schedules.

For a float switch bracket fabrication project, it designed three bracket types (Z-bend rim mounts, discharge standoffs with air gaps, and dual-float series mounts), calculated blank dimensions and bend positions for each tank class, and generated a materials BOM — 32 brackets from two sheets of 1/8″ cast acrylic.

Species Research and Compatibility

Before any animal enters the system, the AI cross-references compatibility against existing tankmates, checks temperature and pH requirements against tank parameters, evaluates niche occupancy (surface, midwater, benthic), and flags territorial conflicts. When a vendor had stream gobies in stock and I needed Stiphodon alternatives, it evaluated each candidate against my specific tank parameters — not generic care sheets, but my actual water temperature, flow rates, and existing community composition.

State Recovery and Data Integrity

The system has survived data corruption incidents where conversation context was lost and files were damaged. The AI rebuilt complete system state from partial backups, conversation history, and cross-referencing multiple file versions — identifying filename/content mismatches, detecting truncated uploads, and reconstructing missing commits from documented deltas.

This is the unsexy but critical capability. A 19-tank system with 200+ animals across dozens of species generates enough state that manual tracking breaks down. The AI maintains referential integrity across the entire dataset.

What This Actually Means

The value isn’t that AI knows about fish. Any aquarium forum can tell you that. The value is that AI can maintain persistent state across a complex system, perform domain-specific calculations on demand, validate data integrity, and execute multi-step operations (fetch → calculate → validate → commit → verify) without dropping context.

I treat my aquarium system the way a process engineer treats a production facility. The tanks are unit operations. The filters are treatment stages. The livestock census is inventory management. The bioload model is capacity planning. And the AI is the operations analyst who never forgets a number, never loses a file, and never gets tired of running the same calculation for the nineteenth tank.

It’s not artificial intelligence managing my aquariums. It’s augmented intelligence — mine, extended by a tool that handles the computational and administrative load so I can focus on the biology.