Your colleague sends you this drawing for review. Would you catch the error?
Version 2.2 of Pen, our AI drawing assistant, is out. It understands those tricky ISO GPS drawings much better now.
Read more about 2.2 and what's wrong in this drawing here: https://t.co/FOrVd7XnSj
@DadaJudith@patrickc@langdock Nice! I find it quite odd to go with a US LLC when the operation is in Europe. Makes many things more complicated and gives you risk exposure to the US jurisdiction. Makes things easier for US VCs of course though.
Pen 2.1 is out. AI drawing review, twice as fast and better at GD&T.
Lots of posts in the last day here that GPT Astra is great at understanding 3D. Our data also show that. We're making use of it and switched to Astra as our default model.
Nothing has changed in terms of our belief in building upon the strongest frontier AI models.
You'll get the breadth of knowledge of the frontier and we don't train on your drawings.
Pen 2 currently runs on GPT 5.6, which shows the best performance in our tests.
Pen 2.0 is live. This is the strongest version of ClearHandoff's AI drawing assistant yet.
It's as fast as its predecessor but catches more mistakes. Pen helps you save human review time and increase the quality of the drawings you release.
With Pen 2.0 we have moved to a new software architecture where multiple AI reviewers each focus on a different error type.
We see that this new architecture allows us to better tackle very complex drawings, where an AI that looks at everything at once can get overwhelmed.
Currently we're looking into specialized subagents.
Early signs are that we can get recall up, but we need to keep false positives in check.
Exciting to cook up new ideas! /3
Our main internal drawing dataset that we optimized AI review against was basically solved since GPT 5.6.
However, for some of our users with complex drawings, we were still performing poorly.
So it was time to upgrade our evals. We put together a harder set of drawings./1
On these drawings, we're down in the 70s percent in our recall metric (the percentage of ground truth errors AI review catches).
And it looks like we need architectural changes, not just tuning, to get back into the 90s.
We're now exploring different architectures./2
AI review for drawings via our API just got a bit simpler.
You can now upload a drawing and trigger an AI review in one call. Previously you needed two.
Here are the full API docs: https://t.co/L0HlcDmqGi
We're launching ClearHandoff on Product Hunt today: AI drawing review for mechanical engineers
If you like what we are building, you can do us a little favor and leave an upvote or a comment
https://t.co/3Gs7OVQkDG
@denisyurchak Check the compensation of CEOs of public companies. Or for startups, check how Austria’s most famous tech guy, Peter Steinberger, made his money.
Our AI drawing assistant now has a name: Pen
Tag @pen on your drawing and ask whatever you'd normally ask a colleague:
* is our laser cutter precise enough for this tolerance?
* anything we can change to reduce manufacturing costs for the part?"
Pen replies in seconds.
@peta2la Yep, I’m also surprised how good the spatial understanding of multimodal LLMs for technical drawings is. Understanding dimensions lines, views, … I’m working with mechanical engineering drawings, not civil as you seem to focus on.