Windyview · applied AI · Ontario, Canada
The farm is the lab.
Windyview is an applied-AI practice run from inside a working mushroom farm. We build perception, telemetry and forecasting systems that operate on-site, on our own hardware — where a wrong answer costs a crop, not a demo.
01 / StatementWhat we hold ourselves to
-
Run where the work happens.
Cameras, controllers and sensors on the local network; models on machines we can touch. Nothing here needs a distant data centre to keep a room alive overnight.
-
Measured, or it didn't happen.
Every model ships behind a pre-registered evaluation: the bar written down before the build, controls that must fail, kill criteria that actually kill. A test a system cannot lose is not a test.
-
Consequences are the curriculum.
Steam, spawn and 90% humidity do not care about benchmarks. A missed flush is lost crop; a wrong forecast is real money. Systems mature here because failure is expensive and immediate.
-
Boring reliability beats a demo.
Heartbeats, ledgers, fail-closed defaults, shadow deployments before live ones. The interesting part of AI is not the launch video — it is week forty, unattended, at three in the morning.
02 / SystemsThree instruments, one discipline
-
eye on-prem perception
A perception engine that turns ordinary cameras into structured events — trays counted, flushes staged, a person entering a room — processed entirely on local hardware, day and night.
The same engine is piloting service perception beyond the farm: an empty teapot, a raised hand, noticed by a single quiet machine in the corner of a restaurant.
live on the farm -
bridge industrial telemetry
The connective layer between decades-old industrial control and modern inference: Siemens-class PLCs, climate sensors and pan-tilt cameras on one supervised local network. Reads are continuous; writes are gated, logged and deliberate.
live on the farm -
harness forecasting & evaluation
Probabilistic forecasting — density estimates and volatility, not point guesses — built walk-forward and evaluated the hard way: against live external scoring, against pre-registered gates, in shadow alongside an incumbent before anything is promoted.
The harness is the product: the same machinery that judges our market models judges our farm models.
competing live
Public standing · third-party scored
Our models compete against every other deployed entry on public leaderboards we do not control or score. Nothing below is self-reported, and both boards are open for anyone to check.
Density forecasting · crypto, equities, commodities
| Horizon | Rank | Score |
|---|---|---|
| 24-hour | 19 / 392 | 0.702 |
| 1-hour | 177 / 392 | 0.494 |
Score is a 7-day rolling rank against the field, where 1.000 is best; 380 of 392 deployed models are currently scoring. Verify →
Structural-break detection · regime change in live series
| Board | Rank | Field |
|---|---|---|
| Overall | 27 | 563 entrants |
The same detector that flags a regime change in a price series flags one in a growing room — a batch turning, a controller drifting. Board moves on submission, not on a clock. Verify →
as of 2026-08-26 08:17 UTC · refreshed every 4 hours
03 / Proving groundA commercial farm, not a sandbox
Windyview Mushroom grows king oyster and gourmet mushrooms year-round: dark incubation, light-tuned fruiting rooms, steel, steam, substrate by the tonne. Every system above earns its keep here first — watching real rooms, reading real controllers, forecasting quantities someone is paid or charged for.
We did not add a farm to an AI company. We added instruments to a farm, and the discipline followed.
Software that has to survive a harvest is different software.
04 / ContactTalk to us
On-site perception, agricultural automation, evaluation infrastructure, or something your operation needs watched carefully — we read everything.
[email protected]© 2026 Windyview · Ontario, Canada · grown & computed on-site
中文 · farm site · AI 快报