On-site AI

AI that runs where the data is

The engine takes the models to the data, not the data to a cloud AI service. They run on hardware at the site, so a field, a clinic, a camera or a warehouse keeps what it collects.

The data stays put

Imagery, records and video are read where they were collected. A cloud AI model would only ever be added by an explicit, disclosed decision.

Work never waits on a connection

The models sit beside the work, so a weak or missing internet link does not stop it.

One model for each job

Each step has its own model, chosen and tuned for that one job. Only the one at work is loaded, so one machine on site runs them all.

Where it stands

Built in two products, the rule for the rest

AgriSmartOne

Built. Seven models read 3 cm drone imagery on hardware at the hub, and no cloud AI is called. The detail is below.

HomeOne

Prototype. The chat that turns a brief into rooms, a plan and a 3D model runs on our own hardware.

The other products

The same rule is the starting point for each of them. Where a product is still in design, its page says so.

In depth: farms

On a farm: reading the land at the hub

AgriSmartOne is where on-site AI has been measured. A disclosed satellite service is the only outside data source.

What the models do

  • Segment (field-segment): cuts every drone tile into regions, about 1.3 seconds a tile.
  • Label (land-label): names each region: crop, bare soil, grass, tree, water or rock.
  • Find by name (object-find): finds huts, tracks, stones and trees from a single word, catching what the first pass missed.

Together they give a whole-field land analysis: tracks, buildings, trees, rocks, crop, bare soil and water, each with its count and area. On DEMO-01, the maize farm in the season film, the AI finds 20 trees, 22 buildings and 0.66 ha of tracks.

18 minto map every tile of a 9.9 ha field, on local hardware
3 cmper pixel: the drone imagery the AI reads
15huts and buildings found on a test tile by the object finder, matching a count by hand
0cloud AI calls. A disclosed satellite service is the only outside data source
Contact

Put the engine to work with us

For partners, investors, government programmes and operators that want to run a pilot. Your message goes straight to the Aionautics team.

Partner with us

Privacy

The contact form stores your name, email, organisation, topic, message and the time you sent it in Aionautics' own database. We use it to reply to you and don't share it.

We measure visits on our own server with no cookies and nothing sent to analytics or ad companies: pages, where you came from, country and city, device, clicks, scrolling, films watched, form fields reached (never what you type) and a masked replay of the page, kept 90 days. Opt out on this browser with this link; details on the privacy page. Fonts load from Google Fonts. To have your message deleted, write to us through the same form.

Film credits

Drone imagery: OpenAerialMap, North East Region survey flown 5 October 2021, CC BY 4.0. Earth imagery: NASA Visible Earth, Blue Marble (public domain). Borders: Natural Earth (public domain). Plants, soil, terrain and the field scenes are generated procedurally and rendered in Blender.

ModelAuthorLicence
Harvester Near Completion (modified)webbsketchesCC BY 4.0
Low Poly Combine HarvestorsiedaltonCC BY 4.0
1975 FUSO Canter T200jinjinjinCC BY 4.0
Isuzu Elf seriesRomanov_13CC BY 4.0
Mado_Ricebags (modified)Korea Heritage ServiceCC BY 4.0
Warehouse FBX Model FreeNicholas-3DCC BY 4.0
Wooden PalletsYadroGamesCC BY 4.0
Microsoft Rocketbox avatars and motion capture (clothing re-dyed)© 2020 MicrosoftMIT