ByteHub AI logoByteHub AI
The platform

Four products. One signed artifact. One policy.

ByteHub turns a model prototype into something a device fleet can run, update, audit, and roll back — on the silicon already in the product, with the cloud used only by policy.

Request product brief See the technology
01 · Edge

EdgeRuntime

Quantized local inference (INT8 / INT4) for low latency, privacy, and offline operation. Execution profiles are tuned per SoC — NPU, GPU, or CPU fallback — so the same signed artifact runs predictably on i.MX95-class, Jetson Orin, and Jetson Thor hardware, with an Android runtime planned for the handhelds and operator terminals that sit beside the fleet.

  • Per-target execution profiles and compatibility rules
  • Runs purpose-trained specialist models — see Models for how a frontier teacher and an open base model become one device’s skill
  • Linux devices today; Android runtime (Snapdragon / Hexagon NPU, LiteRT, ExecuTorch) planned
  • Local policy gate: what may run, what may leave, what needs approval
  • Signature verification on device before any model runs
02 · Operations

ModelOps control plane

Distill, quantize, evaluate, sign, deploy, observe, and roll back model artifacts like firmware. Artifact registry, release management, staged rollout, fleet observability, and an append-only audit log — per tenant.

  • Versioned, signed, target-compatible artifacts
  • Staged rollout with a verified rollback target always present
  • Provenance: which artifact, profile, policy, device, and evaluation produced an outcome
  • Flywheel: device misses become the next training set; the re-distilled student is re-evaluated and re-signed before rollout
03 · Cloud

Policy-gated cloud intelligence

Escalate only approved, minimized context to a frontier model when the task exceeds the device's compute, memory, or connectivity envelope. Default-off. Metadata only unless policy opens the path. Every crossing lands in the audit log.

  • Provider-aware: Vertex AI / Gemini, Amazon Bedrock, Microsoft Foundry
  • Deterministic routing with a logged reason per request
  • Predictable return-to-device behaviour when the link drops
04 · Isolation

Business sandbox

Tenant-isolated context, tools, and audit boundaries for each customer deployment. Business knowledge stays in a sealed boundary — not mixed into a shared context, not used for training, retained only as long as the customer's policy says.

  • One isolated context, key, and retention policy per tenant
  • Approval gates for consequential device actions
  • Tenant-visible audit without prompt bodies in shared logs
What we build on

We do not rebuild the runtime. We make it deployable.

ByteHub adds the production layer on top of the vendor stacks device makers already ship. Names below are the stacks we build on and validate against, not endorsements.

LayerWhat we build onWhat ByteHub adds
Device runtime (NVIDIA)JetPack, CUDA, TensorRT, TensorRT-LLM / TensorRT Edge-LLMSigned model artifacts, per-target execution profiles, verified rollback
Device runtime (NXP)eIQ, Neutron NPU, ONNX Runtime, Yocto BSPSame artifact format and policy gate on low-power industrial silicon
Device runtime (Android, planned)Android NNAPI / LiteRT, Qualcomm AI Engine Direct, ExecuTorchSame artifact and policy gate on rugged handhelds and operator terminals
Device update systemsJetPack OTA, RAUC, SWUpdateModel artifacts get compatibility rules, staged rollout, telemetry, and rollback like OS images
Frontier cloudVertex AI / Gemini, Amazon Bedrock, Microsoft FoundryPolicy-authorized escalation, context minimization, per-tenant isolation, audit
Open runtimesllama.cpp, vLLM, ExecuTorch, LiteRT-LMEvaluated per target; packaged only when measured on our bench
Business model

A software-first control plane for the installed base of Linux edge devices.

Initial buyer: industrial equipment OEMs and fleet operators running Linux-based devices in the field.

Runtime

Edge runtime license

Per-device or per-gateway annual subscription for local runtime, compatibility profiles, policy enforcement, and updates.

Control plane

Enterprise control plane

Per-tenant subscription for artifact registry, release management, audit, fleet observability, and governance.

Usage

Usage + services

Metered cloud intelligence and paid integration or pilot services — designed to become reusable product modules, not recurring custom engineering.

SegmentBuyerPain that creates budget
Industrial equipment OEMsVP Engineering / Head of PlatformFaster, safer release lifecycle across deployed products
Fleet operatorsDirector of device operations / reliabilityLower incident response and field-support burden
Robotics / inspection integratorsCTO / Product engineeringPrivate, offline AI deployment without building the full lifecycle stack
Product brief

Request the product brief.

EdgeRuntime, ModelOps control plane, policy-gated cloud intelligence, and business sandbox — architecture, execution-profile plan, and the design-partner evaluation format.

Request product brief Design-partner inquiry