
Hestus
www.hestus.co/ →Autocomplete for CAD is here
WHAT IT SOLVES
60% of a CAD session is repetitive clicks and menu diving — actual design thinking gets buried
WHY IT'S INTERESTING
No text prompts — it autocompletes natively inside CAD
Instead of a chat box where you describe parts in text, Hestus reads your sketch in real time, predicts your next move, and finishes it with one keystroke. That's AI built for CAD's language — not ChatGPT duct-taped to a 3D viewport
2.5x faster, 4x fewer clicks
These aren't vanity numbers — 'understanding design intent' means training on massive CAD interaction data. Every click eliminated is proof the model actually predicts accurately
Demo shows beta features — public release isn't live yet
The site explicitly says the demo showcases beta features available only to select users — no 'download and go' pretense. Honest framing
「Less clicking. More designing」
TECH GUESS
Likely a plugin architecture for major CAD platforms (SolidWorks/Fusion 360 tier), with a transformer model predicting sketch sequences server-side
DEEP DIVE
The Last-Mile Pain of CAD Interaction
A comment on HN with 224 points captures the problem perfectly. User 2four2, who has used CAD nearly every day for their entire engineering career, wrote: "My day is filled with a tedium of micromanaging Solidworks or NX to simply do what I need. It takes very little time for me to envision a solution, but it takes ages of clunking through UI to actualize it." Hestus's core thesis is that 60% of hardware design time is spent on repetitive clicking and menu navigation. Rather than bolting a chatbot onto CAD, it builds a sketch-sequence predictor that lives directly inside your CAD environment, reads what you're drawing, and offers one-keystroke completion for the next likely action—be it a constraint, dimension, or extrusion.
Autocomplete, Not Chatbot
Founders kevinsane and Sohrab come from system integration and hardware design backgrounds. Their frustration wasn't "AI can't understand my needs" but "why do I have to click so many times to get there." The tagline Less clicking. More designing. is backed by a claimed 2.5x speedup with 4x fewer clicks. Crucially, they chose the word "Autocomplete" rather than "Auto-design." The demo shows basic operations—HV (horizontal/vertical) constraints, equal-diameter circles—which HN user progbits called "trivial and not convincing." But this is a deliberate starting point: eat the most mechanical, repetitive interactions first, then expand.
White-Box CAD vs. Code CAD (OpenSCAD)
An interesting thread in the 87-comment discussion was whether LLMs would revive code-based modeling like OpenSCAD. zackangelo argued it's easier for GPT to generate code than to manipulate GUI state. Robotics engineer TaylorAlexander pushed back: "OpenSCAD just cannot perform appropriately to serve as a primary CAD platform for professional mechanical design." Meanwhile lincoln20xx praised build123d (built on cadquery) for real engineering work, though brew-to-pip installation on Mac M1 was a dead end until conda-forge helped. Hestus took the white-box path—predicting inside the GUI rather than generating code. This is a more conservative but arguably more respectful approach to how CAD engineers actually work.
Honest Limitations and Beta Status
The site clearly states the demo video shows beta features available only to select users; you need to request priority access. kevinsane's pitch was measured: "We hope to create software where it is just easier and faster to prototype and test things in real life." HN user lima pointed out the real hard part—"not backing yourself into a corner with the feature tree over time"—which Hestus's current demo doesn't address. User 2four2 also signaled that SolidWorks support is a prerequisite for adoption, hinting the current build may target only one CAD environment.
Who Should Watch Hestus
If you're a mechanical engineer spending hours daily on sketch-constrain-extrude cycles in SolidWorks or Fusion 360, Hestus deserves a spot on your radar. It's not trying to replace your design judgment—it's trying to collapse those "I know exactly what's next but I still have to click five times" moments into a single keystroke. The real test will be prediction accuracy on complex feature trees and non-standard constraints. For now, it's an early-stage product with clear boundaries, not another AI-everything tool.
Discussion (0)
- No comments yet — be the first.
Related
#121▶ 179AI-made HTML decks you can actually edit — then export to PPTX
#120▶ 113One link to ditch all watermarks
#119▶ 140When's the last time you actually revisited something you saved?