Live colony · listening now

Your hive, right now

source demo baseline node
Listening to the colony…
confidence

The 8 listening bands · 175–525 Hz

The raw acoustic spectrum from the hive mic — eight Goertzel frequency bands. The tallest bar is the colony's dominant tone.

What the shape means

Each band-shape maps onto six possible meanings. The highlighted one is what drove the verdict.

Every number above is a real field from the BeeKeeper engine's /explain endpoint. In DEMO it replays a captured trace; switch to LIVE with the engine running and it polls the colony every few seconds.

Live reasoning trace

How the AI turns a sound into meaning

Eight numbers come off a microphone in a Maine field. Watch them become an interpretation, a verdict, and the cited science that backs it — every value below is a real field from the engine.

strangeness embedding baseline

1 · Signal

Raw acoustic spectrum — 8 Goertzel bands, 175–525 Hz, straight off the hive mic.

interpret

2 · Interpretation

The spectrum's shape maps to six meanings. The one that explains the verdict is highlighted.

classify

3 · Verdict

A frozen classifier — byte-identical to the one running on the Seed — names the state.

ground in research

4 · Research

The verdict is checked against a vector knowledge base of cited bee-acoustics studies — nearest evidence first.

48-hour replay

Two days in the life of Hive #1

A timeline of the colony's state — every reading, every event, scrubbed back hour by hour. This is the dashboard Stuart sees after harvesting the data from the field.

Cited knowledge base · entries

Ask the hive

Everything the system knows about bee acoustics — queenless roars, swarm piping, robbing, thermoregulation — drawn from peer-reviewed studies. Ask a question and hear the sound it describes, synthesized live.

How it works · end to end

A hive that listens, off the grid

No wifi. No cell signal. No mains power. A solar-charged computer sits in a field in Harpswell, Maine, listening to a beehive and deciding — on its own — what the colony is doing.

The chain of custody

1
A custom microphone node listens at the hive entrance.It runs an 8-band FFT over 150–550 Hz — the range where bee sound lives — and sends 48 bytes every ~10 seconds.
2
A solar-powered Raspberry Pi (the "Cognitum Seed") receives those packets over its own little field wifi.No internet. The Seed makes its own network so the mic can reach it.
3
On-device, a frozen acoustic classifier names the state: healthy, chaotic, queenless, swarming, or robbing.The same model byte-for-byte as the one this app shows you under "How it thinks".
4
Results are stored locally on the Seed and checked against a cited knowledge base of bee-acoustics research.Stuart harvests the data when he visits — the whole field procedure needs zero computer.

The honesty model

An AI in a field is only trustworthy if it tells you what it cannot do. This one does:

  • It states its confidence. Every verdict carries a number. Low confidence is shown as low, never hidden.
  • It flags what it cannot hear. The strongest swarm-prediction signal is a ~20 Hz comb vibration — below this mic's floor. The app says so rather than pretending.
  • It labels synthetic vs real. When the data source is a simulation, the badge reads synthetic, not live. No theater.
  • It shows its work. Every verdict links to the peer-reviewed study behind it. You can check the AI.

Why this is hard (and why it matters)

A beekeeper can open a hive maybe once a week, and only on a warm day. A queen can fail, a colony can prepare to swarm, or robbers can collapse a hive in the days between visits — and by the time you look, it's over. A microphone never blinks. By learning each colony's own baseline and listening for the acoustic signatures the research describes, this system can tell a non-technical keeper something they could otherwise only guess at: your hive sounds calm and queenright — or something changed at 3am, go look. That is AI doing real, useful work in a place with no infrastructure at all.

Built for one hive in Maine, designed to honestly report what it knows and what it doesn't.