LAT 24.7956N · LON 120.9967E
NODE: HSINCHU — NTHU
--:--:-- LOCAL
DemoArchitectureSkills PricingTeamGet in Touch
Private Alpha · 2026 · Built at NTHU

EDGEMINDCORE

The operating system for physical AI. Control any robot, sensor, or machine with a single sentence — planned and executed 100% offline, on the device itself. No code. No cloud. No leaks.

0
Cloud calls
<1s
Plan latency*
7+
HW protocols
Composable skills
LOCAL-FIRSTPROMPT-NATIVEHARDWARE-AGNOSTICPRIVACY-PRESERVINGMODULAR SKILLS LOCAL-FIRSTPROMPT-NATIVEHARDWARE-AGNOSTICPRIVACY-PRESERVINGMODULAR SKILLS
00  Live Runtime

SAY IT.
WATCH IT THINK.

This is the EdgeMind runtime, running locally. Pick a command — or write your own — and watch a sentence become a validated, executable physical action plan. Zero cloud calls.

edgemind@local — agent-runtime MODEL qwen-edgeNET OFFLINE ●
// choose a command
// structured output
▸ awaiting command… select a preset on the left, then RUN.
idle · runtime ready
01  The Problem

AI HAS BEEN TRAPPED IN
THE DIGITAL WORLD.
WE'RE LETTING IT OUT.

Today's assistants can talk. They can't do. EdgeMind Core is the universal control layer that gives AI hands — turning a sentence into motion, sensing and decisions across robots, industrial gear, appliances and IoT. No code. No cloud. No limits.

01

Describe

State what you want the world to do, in plain language. No SDK, no wiring diagrams.

02

EdgeMind Plans

The on-device agent parses intent, retrieves Skills, schedules execution and validates safety.

03

Machines Act

The Hardware Abstraction Layer fires the right protocol — UART, ROS, MQTT, Modbus — and it's done.

// 01

Local-First

Inference, planning and memory all run on-device.

// 02

Privacy

Data never leaves the machine.

// 03

HW-Agnostic

From a $4 MCU to a robot fleet.

// 04

Prompt-Native

Language is the only interface.

// 05

Skill Ecosystem

Modular, shareable, composable.

0
Cloud calls required
0
Layer runtime stack
0
Hardware protocols
0
Offline capable
02  Core Architecture

FROM WORDS TO WORLD.

A prompt falls through six layers and comes out the bottom as physical action. Each layer is independently swappable — an Embodied Agent Runtime, not an LLM wrapper.

IN
User Prompt
natural language input
L1
Intent Understanding
extraction · constraints · grounding
L2
Task Planner / Runtime
decomposition · scheduling · recovery
L3
Skill Orchestrator
dynamic skill composition
L4
Hardware Abstraction
unify every protocol
OUT
Physical Devices
robots · sensors · machines
03  Intent Engine

NOT A CHATBOT.

A local LLM parses your words into structured, executable intent: extraction, constraint parsing, temporal reasoning, device grounding and action-dependency analysis. All on the edge.

Llama
Mistral
Gemma
Phi
Qwen
PROMPT → STRUCTURED INTENTlocal.parse()
# "Check the living-room temp. If over 30°C,
#  turn on the fan and notify me."
{
  intent: "environment_control",
  conditions: [
    { sensor: "temperature_sensor",
      operator: ">", value: 30 }
  ],
  actions: [ "turn_on_fan", "send_notification" ]
}
04  Task Planning Engine

THE BRAIN
OF THE BEAST.

The Planner never touches hardware directly. It decomposes the goal, retrieves Skills, schedules execution, tracks state and recovers from failure — a true autonomous loop.

PROMPTwarehouse.patrol()
"Patrol the warehouse. If you hear an
 anomaly, record it and raise an alarm."
Engage mobility module

Begin autonomous patrol route.

Activate audio monitoring

Stream microphone input continuously.

Analyze dB & frequency

Detect abnormal acoustic signatures live.

Trigger camera skill

Capture footage on anomaly detection.

Persist evidence

Store to local memory.

Dispatch alert

Notify the operator immediately.

05  Skill-Based Execution

EVERY MACHINE'S POWER,
PACKAGED.

Each device capability is abstracted into a Skill — metadata, an execution API, safety constraints, resources and capability tags. Load them like ROS nodes, isolate them like containers, call them like AI tools.

skill.move_forward.yamlwheeled_robot
skill_name: move_forward
device: wheeled_robot
inputs: [ distance ]
outputs: [ status ]
permissions: [ motor_access ]
skill.capture_image.yamlusb_camera
skill_name: capture_image
device: usb_camera
inputs: [ resolution ]
outputs: [ image_tensor ]
permissions: [ camera_access ]
// RUNTIME

Skill Runtime

Brings Skills to life — safely.

  • Dynamic loading
  • Permission management
  • Execution sandboxing
  • Timeout & rollback
// COMPOSE

Orchestrator

Skills composed on the fly to satisfy any plan.

  • Retrieve by capability
  • Resource allocation
  • State tracking
  • Dependency resolution
// ECOSYSTEM

Skill Marketplace

A hardware skill economy. Publish once, AI-ify anything.

  • Drone skills
  • Industrial robot skills
  • Smart-home skills
  • CNC & sensor skills
06  Hardware Abstraction Layer

ONE COMMAND.
EVERY PROTOCOL.

AI should never care whether it's talking UART, GPIO, MQTT, ROS or Modbus. Write device.move(x=1.0) once — the HAL translates it to whatever the metal speaks.

Arduino
ESP32
Raspberry Pi
PLC
ROS Robot
CAN Bus
MQTT
Modbus
07  Local-First AI

NO SIGNAL?
STILL ALIVE.

Even with the network down, EdgeMind keeps thinking and acting. On-device inference, offline planning, local vector DB and edge-GPU acceleration — squeezed into low-power silicon with 4/8-bit quantization, KV-cache optimization and speculative decoding.

NVIDIA Jetson
Raspberry Pi
AMD Edge AI
Intel NPU
Apple Silicon
08  Multimodal Sensor Fusion

IT DOESN'T JUST READ.
IT PERCEIVES.

Camera, LiDAR, IMU, audio, depth and temperature streams fuse into a single Physical World State Representation — so the AI reasons about reality, not just text.

CAMERA
HUMAN
detected
MICROPHONE
ABNORMAL
sound event
TEMPERATURE
42°
celsius
LiDAR
0.8m
obstacle

▸ Fused inference: "Possible equipment fault or hazardous event."

09  Safety Layer

THE REAL ENEMY ISN'T
HALLUCINATION.
IT'S PHYSICAL HALLUCINATION.

When an AI moves a motor, a wrong token can break a machine — or a person. EdgeMind wraps every action in hard, non-negotiable guardrails.

// LAYER 1

Permission System

No Skill runs without explicit grants.

  • camera_access
  • motor_control
// LAYER 2

Action Validation

Every high-risk action is pre-checked.

  • Constraint checking
  • Collision checking
  • Resource verification
  • Human confirmation
// LAYER 3

Runtime Sandbox

Malicious Skills are caged. Never reach:

  • Root access
  • Arbitrary device control
  • Memory overwrite
10  Deployment Fields

WHERE IT GETS DEPLOYED.

01

Smart Factory

  • Autonomous inspection
  • Anomaly monitoring
  • Equipment control
  • Predictive maintenance
02

Smart Home

  • NL appliance control
  • Elderly care
  • Security monitoring
03

Education / Maker

  • Zero-code robotics
  • "Follow the red object"
  • Instant control flows
04

Defense / Border

  • No-network ops
  • High-privacy zones
  • Edge deployment
11  The Competition

EVERYONE ELSE STOPS AT TALK.

CapabilityLegacy IoTGeneric AI AgentEDGEMIND
Natural-language controlPartialYesYes
Control physical devicesLimitedRare★ Core
Fully offlineFewAlmost none★ Core
Modular skillsNonePartialFull
Unified multi-hardwareNoneNoneYes
Agent planningNonePartialFull
12  Roadmap

THE LONG GAME.

2026 · Q2 — NOW

Private Alpha

Core runtime, intent engine and HAL on Jetson + Raspberry Pi. First design partners onboarded.

2026 · Q3

Skill SDK & Sandbox

Public Skill definition format, runtime sandboxing and the safety validation suite.

2026 · Q4

Skill Marketplace Beta

Third-party developers publish drone, industrial and smart-home Skills.

2027

Physical AI Infrastructure

Fleet orchestration, cross-device memory and the universal runtime for reality.

13  The Mission

PHONES RUN ON ANDROID.
ROBOTS RUN ON ROS.
REALITY RUNS ON EDGEMIND.

No human should need to understand GPIO, UART, PLC, MQTT or ROS topics ever again. Just describe what you want the world to do.

14  Pricing

START FREE.
SCALE TO A FLEET.

Open-core today, enterprise-ready tomorrow. All tiers run fully offline — you own your data and your hardware.

Open Core
Free
For makers, students and tinkerers. Run it on a single device.
  • Local runtime & planner
  • Up to 3 devices
  • Community Skills
  • Single-node memory
GET STARTED
Studio
$49/node · mo
For teams shipping real products on real hardware.
  • Everything in Open Core
  • Unlimited devices
  • Private Skill registry
  • Sensor fusion & long-term memory
  • Priority support
GET STARTED ▸
Enterprise
Custom
For factories, fleets and mission-critical deployments.
  • Fleet orchestration
  • On-prem & air-gapped
  • Custom HAL drivers
  • SLA & compliance
  • Dedicated engineering
CONTACT SALES
15  FAQ

QUESTIONS,
ANSWERED.

Yes. Inference, planning, memory and the vector database all run on-device. With the network down, EdgeMind keeps thinking and controlling hardware — the offline path is the default, not a fallback.
No. EdgeMind is prompt-native — natural language is the interface. A student can say "follow the red object" and get a working control flow. Developers can still author custom Skills when they want lower-level control.
The Hardware Abstraction Layer unifies Arduino, ESP32, Raspberry Pi, PLCs, ROS robots, CAN bus and MQTT devices. Inference runs on NVIDIA Jetson, Raspberry Pi, AMD Edge AI, Intel NPU or Apple Silicon.
A three-layer safety system: a permission model gating every Skill, action validation (constraint, collision and resource checks plus human confirmation for high-risk moves), and a runtime sandbox that denies root, arbitrary device control and memory access.
Any local LLM — Llama, Mistral, Gemma, Phi or Qwen — optimized for the edge with 4/8-bit quantization, KV-cache optimization and speculative decoding.
Team ALL GOOD, built at National Tsing Hua University in Hsinchu, Taiwan — led by Hector Chiu (CEO) and Yuyi Chang (CTO).

GIVE YOUR MACHINES
A MIND.

See the runtime in action, or get in touch with team ALL GOOD.

16  The Crew

BUILT BY ALL GOOD.

TEAM ALL GOOD · NATIONAL TSING HUA UNIVERSITY
HC
▸ CHIEF EXECUTIVE OFFICER

Hector Chiu

National Tsing Hua University · NTHU
YC
▸ CHIEF TECHNOLOGY OFFICER

Yuyi Chang

National Tsing Hua University · NTHU