Docs

Documentation

Evaluate, install and operate DION AI on your shop floor — architecture, the machining pipeline, knowledge bases, security and licensing.

Overview

DION AI is an on-device CAD/CAM middleware. It reads design intent, reasons over a manufacturing ontology, and proposes verified machining strategies across the systems you already run. All computation happens on the shop PC — your drawings never leave the network.

It is built for shops that run one or more CAD/CAM systems and want to capture machining expertise once and reuse it everywhere — without sending a single drawing to the cloud.

This guide walks through evaluating, installing, licensing and operating DION AI in production.

Quickstart

From zero to a verified program in four steps:

quickstart
# 1. Sign in and download the signed installer
#    Account → Downloads → DION AI (latest)

# 2. Install — registers the add-in + desktop board
DION-AI-Setup.exe /verysilent

# 3. Activate with your license key
dion activate --key XXXX-XXXX-XXXX-XXXX

# 4. Open a part and run the pipeline from the board
#    Board → Pipeline → Run
Tip
Most shops start with a two-week on-premise trial. Contact us and we'll provision it.

Requirements

Windows 10 or 11 (x64), 16 GB RAM minimum (32 GB recommended), and a supported CAD/CAM installation already on the machine.

Local administrator rights are needed for the first install. A few gigabytes of free disk hold the local knowledge bases. A GPU is optional and used only for local acceleration — never required.

Installation

Download the signed installer from your account and run it on the shop PC. The add-in registers with your CAD/CAM host, and the desktop board launches from the system tray. On first run, DION AI provisions its local knowledge bases on the machine.

Note
Installers are code-signed. Windows SmartScreen recognises the publisher — there is no "unknown publisher" warning.

Architecture

DION AI separates a control plane from a data plane so sensitive geometry never has to leave the machine.

Data plane (on-device)
Geometry, feature recognition, ontology reasoning and toolpaths — all computed locally on the shop PC.
Control plane (cloud)
Licensing, software updates and opt-in anonymous telemetry only. No CAD/CAM data is ever transmitted.
The add-in
Bridges DION AI to your CAD/CAM host: it reads the active document and drives the import → recognise → strategy flow in place.
The desktop board
The operations surface — pipelines, the ontology graph, simulation and the agent board, all in one local workspace.
Important
Because sensitive data never reaches the cloud, the attack surface is minimal and compliance is dramatically simpler than with cloud-based CAM.

Knowledge bases

Recommendations are grounded in a set of local, on-device stores. Everything below lives on the machine — nothing here leaves the network.

  • Cutting & process knowledgeCutting coefficients and machining strategies that the reasoning engine draws on.
  • Manufacturing ontologyEntities, constraints and an audit log — the model every recommendation is traced back to.
  • Equipment & toolingMachines, fixtures, holders and setups available on your floor.
  • CAD assets & componentsImported models plus reusable component and part-family libraries.
  • NC program corpusA growing library of verified programs used for case-based recommendations.
  • Documentation indexThe searchable index behind the board's in-app help.
Note
All stores are local SQLite databases, running entirely on-device.

Supported systems

DION AI is vendor-agnostic. It works alongside the major CAD/CAM systems you already operate:

  • SolidWorks
  • Solid Edge
  • Autodesk Inventor
  • Siemens NX
  • CATIA
  • Creo
  • Fusion 360
  • ESPRIT
  • Mastercam

Strategy is expressed against the manufacturing ontology, not a single vendor's API — so the same expertise applies whichever system a job runs in.

Note
All product and company names are trademarks of their respective owners. Listing them indicates compatibility only and implies no sponsorship, affiliation or endorsement.

A custom CNC model

DION AI is building a model purpose-built for CNC — not a general-purpose chatbot adapted to manufacturing. Recommendations stay grounded in the ontology, while the model is trained specifically for machining intent.

A 74-billion-parameter model is currently in an approximately two-month reinforcement-learning program on 8× NVIDIA H100, supported by the NVIDIA Innovation Lab.

Purpose-built
Dedicated to CNC programming, not a general LLM.
74B parameters
Sized for manufacturing-grade reasoning.
Reinforcement learning
An ~2-month RL program, in progress.
8× NVIDIA H100
Trained with NVIDIA Innovation Lab support.
Note
Model architecture, training data and methods are proprietary and intentionally omitted here.

Capabilities

DION AI plans the operations a shop runs every day — across turning and milling, one machine kind per program.

Turning
OD/ID roughing and finishing, facing, grooving, threading, drilling and boring on the lathe (G18).
Milling
Face, pocket and profile milling, plus holemaking — drilling, tapping and boring on the mill (G17).
Feature recognition
Pockets, holes, slots, profiles and turned features resolved directly from geometry.
Posting & macros
Controller-aware NC with Fanuc-style macro support (#variables, M99).
Verification
Backplot and motion review before any program reaches the machine.
Note
Machine kind is locked per NC program — lathe (G18) and mill (G17) cycles never mix in one program.

Security & privacy

DION AI is on-device by design. CAD/CAM data is processed locally and is never transmitted to us — there is no cloud upload of your drawings, features or toolpaths.

The control plane handles only licensing, software updates and opt-in anonymous telemetry. Installers are code-signed, and control-plane secrets are held in a managed key vault.

Because your intellectual property stays inside the factory, the data you most need to protect is simply never in transit.

Note
Found a security concern? Please disclose it responsibly to u2dia@naver.com.

The pipeline

A part moves from import to verified output through a step-gated pipeline. Each step only advances once the previous one returns a verified result, so failures surface immediately instead of propagating downstream.

  1. 01ImportNative CAD or a neutral format (STEP, IGES), read on-device.
  2. 02Feature recognitionGeometry resolved into machinable features.
  3. 03Ontology reasoningConstraint-based strategy — deterministic, not guesswork.
  4. 04StrategyTools, operations and parameters proposed.
  5. 05VerifyHuman approval, with backplot to inspect motion before posting.
  6. 06PostGenerate the controller-ready NC program.
cli
# Run the pipeline on a part — from the board or the local CLI
dion pipeline run --model bracket.step \
  --stages import,recognize,strategy,post

→ state:       Completed
  operations:  12
  output:      verified NC  (pending human approval)
Note
The example above is illustrative. The pipeline runs locally; nothing is uploaded.

Determinism & human-in-the-loop

The pipeline is deterministic: the same model and configuration always produce the same plan. Because reasoning happens inside the ontology, every proposed parameter traces back to the rule that produced it — there is no black box.

Every machining proposal is advisory. A human approves the plan before it is posted or sent to a machine, and each decision is recorded — a transparent, auditable trail of what was proposed and who approved it.

Reasoning & algorithms

DION AI's proposals are not free-form generation. Reasoning is bounded by the manufacturing ontology, and precise values come from deterministic lookups — never from a language model's guess.

Feature → constraint → strategy
Recognised features are matched against ontology constraints — material, machine envelope, available tooling — to select only admissible operations.
Deterministic numerics
Coordinates, tolerances, speeds and feeds are resolved from the ontology and cutting-knowledge stores, so the same input always yields the same numbers.
Bounded search
The engine searches only the space the ontology declares valid, which keeps every proposal physically meaningful and safe.
Full traceability
Each proposed parameter links back to the rule or record that produced it — a reviewer sees why, not just what.
Note
The specific rule sets, heuristics and weightings are proprietary. This page describes the approach, not the implementation.

On-device learning

DION AI gets sharper with use — while every drawing stays on the machine. Learning runs on-device, against your own production history, and is never pooled across customers.

Feature extraction
Parts and operations are reduced to structured feature vectors that capture geometry and process characteristics.
Cohorts
Similar parts and operations are clustered, surfacing part families and outliers across your history.
Case-based recommendation
For a new part, the engine retrieves the closest verified cases and proposes what worked before — a nearest-neighbour style match.
Ensemble of proposers
Several independent strategy proposers are combined by a weighted vote, rather than trusting any single heuristic.
Closed feedback loop
Every approve, adjust or reject is recorded and fed back, so accepted proposals are reinforced and rejected ones are down-weighted over time.
Important
Learning influences ranking and selection only — it never invents geometry and never overrides an ontology constraint. The numbers stay deterministic; the ontology stays the boundary.
Note
Model design, feature sets, weights and training procedures are proprietary and intentionally omitted here.

Licensing & activation

Licensing is per copy — per seat and per machine. You activate with a license key from the board or the CLI.

Activation validates online once, then runs with an offline grace period, so the shop floor keeps working without a constant connection. Re-validation happens automatically the next time the machine is online.

Licenses are tied to your deployment; there is no self-serve checkout. Every engagement begins with a conversation and a tailored quote.

Deployment & updates

DION AI is deployed on-premise / on-device. It installs per machine and runs entirely inside your network.

Updates improve the ontology, vendor coverage and functionality. They are delivered through the control plane and governed by the license agreement unless separate terms apply. You stay in control of when updates are applied.

Results

Validated across three pilot customers, each running DION AI on their own parts during a two-week evaluation.

+20%
Production efficiency
+30%
On-time delivery
−60%
Programming time
Note
Outcomes vary by part mix and shop.

Roadmap

  • CNC-specialized model — the 74B model now in RL training moves into the reasoning loop.
  • Wider coverage — more CAD/CAM hosts and controller posts.
  • Deeper on-device learning — richer cohorts and case-based recommendations as your history grows.
  • Broader operations — continued breadth across turning, milling and holemaking.
Note
Directional, not a commitment; sequencing may change.

Glossary

On-device
All computation runs on the shop PC; data does not leave the machine or your network.
Control plane / data plane
The control plane handles licensing and updates in the cloud; the data plane handles geometry and toolpaths locally.
Ontology
A constrained knowledge model of machines, tools, features and the relationships between them, within which DION AI reasons.
Feature recognition
Resolving raw geometry into machinable features such as pockets, holes and profiles.
Backplot
A visual replay of a toolpath used to verify motion before the program reaches the controller.
Human-in-the-loop (HITL)
A required human approval step before any machining output is posted or executed.
Post / post-processing
Generating the controller-specific NC program from a verified strategy.
Determinism
The property that identical inputs and configuration always yield the identical plan.
Cohort
A cluster of similar parts or operations, used to surface families and outliers and to drive case-based recommendations.
Ensemble
A combination of several independent proposers whose outputs are merged by a weighted vote, rather than relying on one.

Support

Enterprise customers get a direct line to the engineering team — not a ticket queue.

For evaluations and quotes, contact us and we'll schedule a two-week on-premise trial on your own parts.

Troubleshooting

The add-in doesn't appear in my CAD/CAM host

Confirm the host version is supported and re-run the installer with administrator rights. The add-in registers during install.

Activation fails while offline

DION AI keeps working through its offline grace period. Re-validation runs automatically the next time the machine is online.

The desktop board isn't in the system tray

Relaunch it from the Start menu. It runs locally and does not require a network connection to open.

FAQ

Where does my data go?

Nowhere. DION AI runs on-device; your CAD/CAM data stays inside your factory.

Which systems are supported?

Every major CAD/CAM system you already run — strategy is vendor-agnostic.

Do I need a constant connection?

No. Activation validates online once, then runs with an offline grace period.

Can outputs run without review?

No. Every machining output requires human approval before it is posted or executed.

Can it handle both lathe and mill?

Yes — but one machine kind per NC program; lathe (G18) and mill (G17) cycles never mix.

Is the AI model required to run?

No. Reasoning is ontology-bounded and deterministic; the custom model augments it, it does not replace the deterministic core.