Private edge AI for on-site intelligence.
Polyhedrontech helps organisations run AI workflows locally — where data privacy, real-time response, power budget, and deployment simplicity matter. One package: edge appliance, hardware-aware runtime, low-code studio.
One private-AI deployment package.
Three integrated layers, purpose-built to move AI from lab prototype to secure on-site pilot — with less infrastructure burden.
Edge AI Appliance
A compact device deployed on-site, designed for offline or private-network environments. Purpose-built for a small edge footprint — labs, factories, classrooms, vehicles.
Optimised AI Runtime
Hardware-aware execution layer for model loading, quantisation, caching and CPU/NPU orchestration. Squeezes real inference performance out of edge silicon.
Low-Code AI Studio
Browser-based interface for configuring models, APIs, and workflows — without firmware-level engineering. Turns approved AI recipes into repeatable deployments.
Built for on-site AI constraints.
We prioritise scenarios where privacy, latency, and integration effort actually block deployment — and where pilots produce measurable outcomes.
Industrial & robotics
Edge intelligence for robot arms, machine vision, local decision support, and factory training environments.
Pilot metrics: latency · accuracy · local uptime · integration time
Education & research labs
Private AI for teaching, simulation, experimentation, and campus-controlled model deployment.
Pilot metrics: setup time · student access · workload coverage
Regulated enterprise AI
Local assistants for document workflows, R&D knowledge, audit-sensitive data, and air-gapped environments.
Pilot metrics: data-control fit · task completion · user adoption
Runtime engineering that makes local AI viable.
We split LLM inference into fine-grained work units and place each unit on the hardware best suited for it — dynamically, per stage, per device.
Neuron-cluster decomposition
Inference splits into fine-grained clusters. Dense clusters run on the NPU, sparse ones fall back to the CPU — dynamically, per inference stage.
Segmented neuron cache
Hot neurons stay in fast memory to avoid repeated I/O. Loading and computation overlap so I/O latency is masked instead of exposed.
Hardware-aware planner
Before runtime, we profile CPU / NPU / memory / storage and generate an execution plan tailored to the device — cache sizes, placement policy, quantisation levels.
The result: privacy stays intact (no cloud round-trip), latency stays low (inference sits next to the data), and power stays within edge budgets.
A spin-off of Hong Kong Industrial AI & Robotics Centre.
Polyhedrontech is currently in the HKSTP Incubation Programme, building an integrated private-AI deployment package: appliance, optimised runtime, and low-code studio.
We focus on the operational constraints that stop enterprise AI from moving beyond demos — privacy, latency, power, and integration effort — and package the answer as a repeatable pilot kit.
Explore a pilot.
Running AI locally under privacy, latency, or integration constraints? Tell us what you're deploying — we'll follow up within one business day.
Prefer email? Reach the founder directly at shawnqi@polyhedrontech.com.