Engineering Intent Model

The most accurate model for reading engineering drawings.

LevelPlane reads function, fit, and criticality from 2D mechanical part drawings across automotive and aerospace, and runs sourcing, supplier matching, and risk on top of it. Benchmarked against frontier vision-language models.

Live with OEMs and suppliers across automotive and aerospace Singapore · Kochi · San Francisco · UK · Mexico

Bearing assembly · BRG-3214-A

STEEL 52100 SCALE 2:1

Benchmarks

Highest fidelity at every stage of the read.

Evaluated against the latest frontier vision-language models including GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro on intent extraction from 2D engineering drawings. Full benchmark releasing August 2026 against a held-out test set of 250 drawings sampled across automotive and aerospace, scored by domain experts.

Releasing August 2026

Test set

250 drawings, automotive and aerospace, held out

Evaluation

Per-feature accuracy, scored by domain experts

Compared against

GPT-5.5, Claude Opus 4.7, Gemini 3.1 Pro

Status

Methodology and results forthcoming

Try it yourself EIM · live

Watch the read.

Drawing arrives. EIM perceives geometry, resolves structure, infers intent, then surfaces what matters for sourcing. The full reasoning chain, in fourteen seconds.

EIM · live

v0.4

Engineering intent

Bore · Ø 52 H7

Bearing seat

Precision-ground · class A

Perpendicularity ⊥ 0.02

Mating surface

Drives CMM inspection

Surface finish · Ra 0.8

Class B sealing

Milling, no secondary

Sourcing intelligence

Should-cost

$48.20 / unit

12k units · 4-axis machining

↗

Suppliers matched

3 ranked

Class A surface · NADCAP

Risk flag

Tolerance stack on bore

Review with engineering

Architecture

Not one model. The right model for each layer of the read.

EIM orchestrates frontier vision-language models, graph reasoning over part structure, and large language models for intent. Composed and verified as one system. We do not compete with the labs. We use them, and beat them by composing them with engineering ground truth.

L1 / Perception What it sees

Detects views, primitives, annotations, dimension lines, GD&T frames, surface finish symbols, and title blocks. Frontier VLMs fine-tuned on engineering drawing primitives.

Claude Sonnet 4.6 Gemini 3.1 Pro Custom fine-tunes

L2 / Structure How it holds together

Resolves dimension chains, datum relationships, GD&T composite references, and geometric tolerance stacks. The parsed drawing is mapped to a structured graph and reasoned over symbolically.

Graph representation Symbolic constraint solver

L3 / Intent Why it matters

Infers function, criticality, sourcing implication, and design intent from the structured representation. Frontier LLMs conditioned on L1 and L2 outputs, retrieval-augmented with your historical context.

Claude Sonnet 4.6 Gemini 3.1 Pro Customer-grounded RAG

Orchestration Routing & verification

Routing, evaluation, fallbacks, and ground-truth checks across all three layers. This is where most of our IP sits. Not in any single model, but in the system that picks the right one for each sub-task and verifies the answer against engineering ground truth.

Why orchestration over a monolith. No single frontier model wins across perception, structure, and intent. By composing the best model for each sub-task, and replacing components as the labs ship, EIM compounds every advance in frontier AI. Every Claude or Gemini release makes EIM better automatically.

Built on top of EIM

Every workflow inherits the read.

The same model that reads the drawing decides what to source, who to ask, what to compare, and what to flag.

Two sides, one model

Whether you buy parts or make them, the read is the same.

For OEMs

Should-cost every part. Benchmark every supplier. In hours, not weeks.

The analysis your sourcing team would do if they had time, run automatically the moment a drawing lands. Should-cost, supplier benchmarking, and risk surfaced from the drawing itself.

For suppliers

Quote like a team three times your size.

RFQ ingest, structured response, sub-supplier dispatch. First quote back wins.

Research / 2026

A drawing is a record of decisions. We are building the system that reads them.

EIM is a research program, not a feature.

We're building a model that reads engineering drawings, orchestrating frontier vision-language models, graph reasoning over part structure, and large language models for intent inference. Starting with 2D drawings, expanding to 3D CAD, statements of requirements, and supplier specifications.

Two patents filed on contract-native agent authority and outcome-gated control.

Pricing

Simple, usage-based pricing.

Start free. Pay as you go. No subscriptions, no seats.

Free trial — no credit card required.
Try two drawings for free. See exactly what EIM reads before spending a cent.

Starter

$199

100 credits

$1.99 / credit

Professional

$899

500 credits

$1.80 / credit

Scale

$2,999

2,000 credits

$1.50 / credit

How many credits does a drawing use?

Simple drawing

~20 credits

Basic part, few features, standard tolerances

Medium drawing

~40 credits

Multiple views, GD&T, several annotations

Complex drawing

~80 credits

Assembly, dense tolerancing, full title block

Need more volume or custom integrations?

Move to Platform — white-glove onboarding, volume pricing, dedicated support, and SLA. Designed for OEMs and Tier 1 suppliers running at scale.