Start with the hardware already there
A camera, microphone, GPS point, local file, or calendar can remove a step from the user’s day. It should earn its place in the workflow.
How Helio runs AI
Put the model close to the job.
Helio builds AI for tasks that start with a photo, a recording, a location, or a file on the user’s machine. The model may run on the device, an edge worker, a private network, or a connected service. We choose the boundary from the job’s latency, privacy, and connectivity.
Deployment boundaries
There is no single place for every model.
We place each step by asking what the device can see, how quickly the user needs an answer, what must stay local, and what can wait for a connection.
| Product or workflow | Device capability | Runs close to | Why |
|---|---|---|---|
| PlantCare | Camera + local model work | iPhone or iPad | The user gets photo feedback without sending every image away. |
| Safari Go | Camera + offline content packs | On-device identification | A wildlife check is still useful when the network is not. |
| VoiceAgenda | Microphone + Apple device services | Supported iPhone or iPad | The recording and the first pass stay with the person who made them. |
| Farmbit Mobile | GPS + media + offline queue | Phone capture, connected sync | Field crews need to finish the visit before the network catches up. |
| Business workflows | Documents + private context + integrations | Edge worker, private network, or cloud where appropriate | Different steps can use different runtimes instead of forcing the whole workflow into one cloud. |
How we make the call
Device-native is a practical choice.
Keep the immediate step close
When latency, privacy, or connectivity matters, capture, retrieval, or inference stays on the device or nearby edge runtime.
Make remote calls a choice
If a larger model or connected service helps, call it for a reason. The product should show that handoff and what it sends.
Give the user a way out
A useful system can say “I do not know”, queue the work, ask for more information, or hand the decision back to a person.
When local inference is the wrong answer
Some jobs need a connection or a specialist.
A connected system, a larger remote model, or a human specialist may be the better choice. We keep that step explicit, log the handoff, and give the user a safe fallback.
For a real workflow
Show us the device, data, and handoff.
We will identify the smallest deployment boundary worth testing.