Vision
Local recognition and detection from camera input.
EDGE AI PRODUCT ENGINEERING
We build AI that runs on real devices.
Helio deploys vision, speech, and language models across mobile, desktop, and embedded devices—then builds the production software around them.
For product teams · For individual products
iOS · Android · macOS · Embedded Linux · Rockchip
VOICEAGENDA / ON-DEVICE SPEECH V1
Spoken input
Apple Speech where available
Foundation Models where required
Transcript · intent · reviewable actions
Runs on supported Apple devices
No third-party speech or LLM service in the V1 workflow.
Device evidence
AI where the data is created.
Local inference can reduce cloud dependency, support disconnected operation, and give a product a clearer data boundary. The trade-offs depend on the model, hardware, and architecture.
Speech
Transcription and intent from an intentional recording.
Language
Extraction and annotation close to the working document.
Devices
Mobile, desktop, embedded, and nearby edge runtimes.
What we build
Three ways we put the model close to the work.
Every engagement starts with the task and the device boundary, then carries the useful result through the product around it.
Embedded computer vision
Deploy camera-based recognition and classification on constrained hardware and embedded Linux devices.
On-device applications
Build native mobile and desktop products around locally executed vision, speech, and language models.
ModelInput
RuntimeReview
ProductAction
Model-to-production integration
Connect inference to storage, review, synchronization, telemetry, updates, and the product experience required in production.
Product evidence
Edge AI, shipped into products.
Our work spans speech, language, vision, mobile applications, desktop software, and embedded targets. The facts below are separated from results we have not yet approved for publication.
Open the app portfolio Embedded vision deployment Technical evidence to confirm
Anonymous technical case · Evidence to confirm
Embedded vision deployment
A camera-led inference path for constrained hardware. The public hardware, runtime, task, and benchmark details are intentionally held until they are approved.
- Target hardware
- Rockchip / board to confirm
- Runtime
- Embedded Linux / runtime to confirm
- Task
- Camera task to confirm
- Input
- Camera source to confirm
- Deployment
- Embedded target to confirm
- Result
- Metric to confirm
Public product
Safari Go
Local visual recognition for a focused wildlife product, with a private field journal and installed destination content for offline use.
- Platform
- iPhone
- Task
- Supported wildlife identification
- Inference
- On device
- Boundary
- Installed destination content works offline
Public product
VoiceAgenda
A spoken thought becomes a Daily Note and reviewable actions while the product keeps the user in control of what reaches Calendar or Reminders.
- Platform
- iPhone and iPad
- Input
- Voice
- AI tasks
- Transcription · intent extraction
- Runtime
- Apple Speech · Foundation Models where required
- Data boundary
- No developer-hosted content service; review before actions
Public product
WolfeWriter
A native Mac writing workspace that keeps manuscript context, Story Memory, and proposed AI changes close to the project.
- Platform
- macOS
- Input
- Manuscript
- AI tasks
- Extraction · annotation · reviewable suggestions
- Runtime
- Foundation Models · Core ML where available
- Storage
- Local-first
Public product
PlantCare
Photo-guided pruning help at the point of capture, with a defined plant set and explicit limits around unsafe work.
- Platform
- iPhone and iPad
- Input
- Plant photos
- AI task
- Supported plant and pruning guidance
- Runtime
- Local model work · detail to confirm
- Boundary
- Cautious guidance and human judgment
Delivery chain
From model to deployed device.
Feasibility, adaptation, product work, and validation are one delivery chain—not four disconnected demos.
Model and task Target hardware Optimized inference Production product Device validation
- 01
Prove device feasibility
Evaluate compatibility, accuracy, latency, memory, throughput, power, and the constraints of the target hardware.
Accuracy · latency · memory - 02
Adapt the inference pipeline
Convert, quantize, replace unsupported operations, and integrate the native or embedded runtime that fits the device.
Conversion · quantization · runtime - 03
Build the production product
Add input pipelines, local storage, review, synchronization, failure handling, telemetry, and the user experience around the model.
Storage · review · sync - 04
Validate on real hardware
Test sustained performance, disconnected behaviour, thermal stability, model quality, and release conditions on representative devices.
Thermal · offline · release
Data boundary
Local by design. Connected when justified.
The device can be the primary execution environment. Cloud services can be added for synchronization, fleet management, collaboration, or work that is not practical locally.
Camera, voice
or documents Local inference Local product
workflow
or documents Local inference Local product
workflow
Optional sync Cloud services
We choose the boundary from the job, not from a default architecture.
- Raw inputs can remain on the device when the product requires it.
- Cloud inference is an explicit architecture choice, not a default dependency.
- Outputs, sources, model versions, and review state can remain traceable.
Release record
The device is part of the release
A release record starts with five fields: task, input, target device, runtime, and failure state. A model result is only meaningful beside the device that produced it. A vision model headed for an embedded Linux board is a different release than one headed for a phone.
VoiceAgenda V1 uses an explicit recording, Apple Speech where available, and a Daily Note before a Calendar or Reminders change is confirmed. Conversion, quantization, unsupported operations, runtime integration, and hardware profiling belong in the same record as the artifact. Helio does not publish a latency, accuracy, or memory figure on this page without a named device, input set, and repeatable command.
VoiceAgenda speech boundaryField notes
Notes from building at the edge.
Qualified starting point
Have a model, a device, or a workload that should run locally?
Tell us what the system needs to recognize, understand, or generate—and the hardware, connectivity, latency, privacy, or power constraints it must operate within.