AI + Biology · Field note
The cloud is someone else's computer, on someone else's network, with someone else's uptime. For a living culture, that is three reasons too many.
Ten years ago, "AI" in a product pitch almost always meant the cloud: sensors uploaded, a server reasoned, results came back. For biological cultivation that architecture fails on three axes at once. Latency: a condensation event does not wait for a round trip. Reliability: a culture must be controlled at 2 a.m. during an ISP outage. Privacy: a lab's grow data is its research capital, and uploading it by default is a decision, not a feature. Edge AI — inference on the device itself — is how a closed loop stays closed.
Consider what a single control decision costs in each architecture. Cloud: sample the sensor, publish over MQTT, broker routes, server infers, command routes back, actuator applies — a chain of network hops where any one failure turns the loop into an open one. Edge: sample, infer on the microcontroller in milliseconds, apply. The difference is not speed on a good day; it is behavior on a bad one. When the network dies, a cloud controller is blind and mute; an edge controller does not notice, because the network was never in the loop.
The hardware reality is friendlier than the marketing suggests. A quantized digital-twin model for a chamber or reactor is a few hundred kilobytes — it fits comfortably on an RP2040 or ESP32-class board, and the fastest path to the camera-based vision module is a small USB accelerator that costs about as much as a good dinner. The models are trained on a workstation and distilled down to what the device needs; the device never pretends to be a training rig.
Privacy on the edge is structural rather than aspirational. In Mycophyte, camera frames are processed on-device and never leave it by default; in AlgaePhyte, sensor logs stay local, and what the federated network receives is the twin's tuned parameters, bounds-checked and signed. Sharing becomes an explicit act instead of an architectural accident — a posture that matters to a lab protecting protocols and to a hobbyist who simply doesn't want their kitchen photographed.
Edge does not mean isolated. Devices still publish telemetry and join federated rounds over MQTT when they choose to; the edge is what keeps those joins optional. The design principle is worth stating plainly: the network should enhance a controller that already works, never be the reason it works. That single sentence explains the failsafe architecture — deterministic local rules underneath everything — and it is the answer we give anyone who asks why the intelligence lives in the box.
For a deeper look at the trade-offs, the docs cover the inference stack and MQTT topics, and the federated learning article covers what happens when edge devices choose to learn from each other.