Every technical drawing is out of date the moment it's drawn

Reading every callout

The first thing we do is read a drawing the way an experienced engineer does, but exhaustively, and at scale. Through our own, in-house Machine Learning models we extract every callout on the sheet and turn it into structured data: a Ø12 H7 bore, a flatness control of 0.05, a position tolerance of Ø0.1 referenced to datums A, B, and C, an Ra 1.6 surface finish, a "break all sharp edges" note tucked in the corner. Each becomes a machine-readable object rather than raw pixels.

That sounds mundane until you realize that no downstream system can do this today. The drawing is the authoritative manufacturing document, and it's the one artifact in the pipeline that software treats as opaque. Once the callouts are structured, the world is your oyster: everything else becomes possible.

Revisions are a record of decisions

Drawings never exist in isolation. Rev A becomes Rev B becomes Rev C, and most teams track that as little more than a date and a one-line note in the title block: "updated per ECN 1043." Our Drawing Intelligence layer compares revisions at the level of the callouts themselves. We can tell you that between Rev B and Rev C a bore tolerance tightened from ±0.1 to ±0.02, that a chamfer was added, that a material note changed from one alloy to another.

Each of those diffs is a decision. And once you can see the decisions, you can reconstruct the chain of them: why a part ended up the way it did, where tolerances crept tighter over time and what prompted it. For every part with a revision history, we can recover the design narrative that produced it; a narrative that often lives nowhere except in the heads of the people who were in the room at the time.

Standards as context

Callouts mean nothing in isolation; they live inside a web of standards: ISO 1302 for surface texture, ISO 1101 and ASME Y14.5 for geometric tolerancing, and a long tail of company and industry specs on top. Because we read callouts as structured data, we can cross-reference them against those standards automatically: interpreting what a symbol formally requires, flagging a tolerance that's impractical for the called-out process, catching a surface finish the chosen manufacturing method can't actually hold.

At that point the drawing stops being a static sheet of A0 paper and becomes something that can be checked, questioned, and augmented. We surface design and manufacturing concerns on the drawing itself, grounded in the standards the drawing is supposed to be following.

The heuristics an engineering company doesn't know it has

This is where it gets interesting at volume. Many engineering organization have documented internal standards: design guides, tolerance tables, approved-process lists. And every engineering organization also runs on a much larger body of undocumented rules: the things experienced engineers simply do, the defaults they reach for, the limits they never exceed even though no spec says so. That tacit knowledge is the real operating system of a design team, and it normally walks out the door when senior people leave.

Read enough of a company's drawings and those heuristics stop being invisible. The patterns recur across parts, across engineers, across years. A particular class of feature is never specified tighter than a certain bound. A certain relief is always added in a certain place. A finish is always chosen for a certain kind of mating surface. Our Drawing Intelligence layer can distill these de facto rules from the drawings themselves, turning what was implicit and unwritten into something explicit and queryable. Not the standards a company says it follows, but the ones it actually does.

One part, two representations

The 3D model and the 2D drawing each know something the other doesn't. A CAD model carries deep geometric and topological understanding: faces, edges, how features adjoin, how the whole thing fits together in space, but it says almost nothing about how the part should be manufactured or inspected. The drawing carries exactly that manufacturing and inspection intent, but it discards most of the rich 3D structure in the act of flattening the part onto a page.

The next step is a shared latent space where 3D models and 2D drawings align, so each representation can inform the other. The model's geometric and topological understanding onto the drawing, so the system knows which face an Ra 1.6 callout actually refers to, not just where the symbol sits on the sheet. And the drawing's manufacturing details project back onto the model, so the 3D geometry carries the tolerances, finishes, and notes that were only ever written on paper. Two views of one part, finally speaking the same language.

Toward a single source of truth

All of this points at one goal. Today a drawing forks from the model the moment it's created, and from then on the two drift apart, reconciled only by manual effort and good memory. We want to replace that fork with a living link: edit the 3D model, and the drawing updates to reflect it. Once that's in place, the drawing stops being a frozen snapshot and becomes a view that stays current. A single source of truth that the whole team can trust.

Getting there means teaching machines to read drawings as fluently as engineers do, to understand revisions as decisions, to hold a company's real heuristics, and to see the model and the drawing as one object rather than two. That's the work. It's also, we think, the foundation for inventing better hardware, together.

If you’re interested in learning more about our research, please reach out to contact@getencube.com, or jobs@getencube.com, if you want to become part of pushing the frontier of mechanical design engineering.