TechNode · AI Software

AI that connects
people and machines

TechNode develops AI software that helps people work with hardware and systems more easily. Our AI interprets intent and context from video, voice, and text, translating them into actionable instructions for customers’ equipment and operational systems.

Partnership & Deployment Inquiries →Investment Inquiries

SORIMultimodal Intelligence for the Field

SORI

Human intent,
translated into action

AI software that understands instructions and field information, and passes only validated tasks to machines. High-risk actions require final human approval.

SORI Performance Index

SORI-1 development targets · 2027

Command Accuracy

95%

Voice commands with target, action and value all correct

Number·Callsign Error

≤ 1%

Recognition errors on altitude, heading, callsigns and other numbers

Unsafe Command Block

99%

Rule-defined unsafe or forbidden commands stopped before execution

P95 Latency

< 1.0s

95th-percentile time from end of speech to command delivery

Edge Footprint

≤ 3GB

Total memory on an 8GB-class embedded GPU

These are SORI-1 development targets, not measured results. Use on real equipment requires separate safety and field validation.

Current internal tests

Measured · voice and text commands

Unsafe Pass-through

0/500

Commands that should be refused yet passed rule validation

Validation Latency

131ms

Mean latency of the rule-validation stage

Output Validity

100%

Valid JSON output format (1,000 runs)

Limited internal tests, separate from safety validation on real equipment. 100% is output-format validity, not reasoning accuracy.

How SORI works

How a human instruction becomes
a verified task

01
InstructionA task requested by voice or text
02
IntentTask, target, quantity and destination extracted
03
Pre-execution checkAllowed commands, values and conditions verified
04
ExecutionTask sent to the device or control system
05
ConfirmationStatus returned and confirmed by the operator

Conditions met: executeUncertain: ask againOut of scope: refuseHigh risk: human approval

Industries

Expanding across defense, manufacturing and logistics

DEFENSE

Defense · Aviation

Accurate decisions and safe execution in complex situations

Understanding instructions for aviation and unmanned systems, supporting operator decisions

MANUFACTURING

Manufacturing

Flexible work that understands what people mean

AI deployed for the factory floor, linked to equipment and business systems

LOGISTICS

Logistics

Stable operation that adapts to real-time change

Normalizing orders and work instructions, linked to logistics systems and robots

Projects

Applying SORI across different fields

Hanwha Systems In progress

Challenge: Interpret complex instructions and execution conditions, and verify the allowed envelope independently.
TechNode's role: Turn operational instructions into structured commands and verify execution conditions with independent rules.
Deliverables (target): Instruction interpretation and command verification prototype
Test results per operational scenario

LG AI Research In progress

Challenge: Run AI reliably on constrained computing hardware.
TechNode's role: Compress the EXAONE model and verify runtime performance per deployment environment.
Deliverables (target): Compressed model and runtime package
Performance comparison per environment

Kakao Mobility In progress

Challenge: Turn worker instructions into robot tasks and integrate with the control system.
TechNode's role: Structure spoken instructions into task data and send commands to WMS and control APIs.
Deliverables (target): Natural-language task module
Robot movement demo linked to the control system

Core Capabilities

Built on domain expertise
and continuous validation

From virtual scenarios to field tests — SORI simulator situation board during a barometric altimeter fault (synthetic data)

Validated commands Built

Constrained decoding allows only permitted command fields, and a rule engine separate from the model cross-checks them against sensors and conditions. High-risk commands need final operator approval.

Proven in simulation first Built

We fly a virtual aircraft through to landing on the AI's decision and judge the outcome in code. 52 test sets cover every pair of conditions across 86,400 valid combinations of 8 condition axes.

Edge-ready models In development

Fine-tuning, knowledge distillation and 4-/3-bit quantization of a compact sLLM to run on 8GB-class embedded GPUs.

AI that connectspeople and machines

Human intent,translated into action