Semiconductor Tooling Briefing: Open Mesh-to-CAD Reconstruction Meets the Agentic EDA Wave

Two separate items landed in the same Engineering Paper cycle, and for semiconductor, board and system teams they read as one signal. On the geometry side, scan and print data that previously had no editable history is now being reverse-engineered into parametric construction sequences. On the flow side, architecture planning, digital and analog execution, PCB verification and tool orchestration are being handed to AI agents that operate existing engineering tools under human direction. The useful question for an engineering or procurement group is therefore not whether these capabilities exist, but where they can be inserted today and where their stated limits begin.

At a glance: which layer of the design chain moved

Item	Segment addressed	Change described in the source material
CADFit	Mechanical/CAD geometry capture	Open-source mesh-to-CAD reconstruction that emits editable construction sequences
Celus Design Platform	Electronics design planning (EDA)	Agentic workflow available in beta, positioned ahead of schematic capture
ChipAgents	Semiconductor design software	Agentic engineering platform expanded across digital and analog chip design
Circuit Mind AVE	ECAD / PCB verification	Agentic Verification Engine that checks schematics against reference documents and internal checklists
Cadence Maestro	Chip and system design (EDA/SDA)	Portfolio of domain-specific AI models used by AI Super Agents to drive Cadence tools
Guest analysis	Engineering.com / Luminary	Distinction drawn between physical AI and physics AI

CADFit: mesh data in, editable parametric history out

CADFit is not the first attempt at this class of problem; earlier editions of Engineering Paper referenced it as a mesh-to-CAD tool in the vein of Backflip AI and Autodesk Fusion’s AutoTimeline. What separates the project is its availability model and its output form. CADFit is open-source and can be freely used for non-commercial work. The initial research was funded by Altair, now part of Siemens, and Ghadi Nehme — a PhD student at MIT who leads the project — together with Siemens have filed for a patent for the CADFit method.

The underlying technique is defined formally in Nehme’s paper, which describes the method as a “hybrid optimization-based CAD reconstruction framework that recovers complex, editable CAD construction sequences from meshes by incrementally fitting and validating parametric operations using geometric feedback.”

Put less academically, “[CADFit] basically starts with the mesh and then extracts some key sketches, similar to how engineers would extract them,” and those sketches are then converted into solids, consolidated with unions and cuts, compared back against the initial mesh, and re-iterated until the gaps are filled.

A video narrated by Nehme walks through that loop. According to Nehme, a simple machine-learning model identifies which sketches are meaningful; it is not essential to the framework, but it speeds the process up by a factor of three.

Current operation coverage and the geometry it cannot yet solve

On capability, Nehme is explicit about the boundary: “Currently, the operations that CADFit supports are extrusions, revolutions, fillets, and chamfers, and then also union and cut and all the sketch operations,” he said, before adding the constraint that “But if a shape is very complex, like loft-based or also sweep-based shapes… the math will struggle.” He added that he is working on a new CADFit version aimed at handling such shapes better.

Output format, viewer and integration status

The models produced by CADFit are expressed in CadQuery, an open-source Python library, which matters for teams that script geometry rather than click through a GUI. Nehme has also built Studio CAD, his own viewer intended for “human friendly editing of the shapes.” There are no plugins for common CAD software yet, but Nehme is working on one for Siemens NX.

Roadmap claims to hold the vendor to

“The future iteration of this work will potentially be able to be really robust on a very huge set of shapes,” Nehme said, while arguing that “But this one is already pushing the boundary by a huge amount… This method is able to really extract the key components of the geometry much better than what’s been seen before.”

The agentic layer: four vendor moves around a single theme

EDA developer Celus has launched an agentic workflow for its Celus Design Platform, available in beta. The agentic design assistant “helps hardware engineers plan the architecture and component work at the start of an electronics design, before schematic capture, and carries out that plan once the engineer approves it,” according to the company’s press release.

ChipAgents, a semiconductor design software developer, said it is expanding its agentic engineering platform across digital and analog chip design. “ChipAgents’ full-stack approach enables engineers to define objectives, constraints and critical judgment calls while AI agents carry more of the execution forward, reasoning over design context, operating engineering tools, evaluating results and determining the next action toward closure,” reads the press release.

ECAD developer Circuit Mind has launched AVE, or Agentic Verification Engine. The company calls it an “AI application that checks Printed Circuit Board (PCB) schematics against component datasheets, application notes, and each organization’s own design review checklists, catching errors that slip past even careful manual review before they turn into costly board respins.”

EDA developer Cadence has launched Maestro, a portfolio of domain-specific AI models for agentic chip and system design. “Combining our Maestro portfolio of domain-specialized models with leading frontier and open-weight models enables our AI Super Agents to drive Cadence EDA and SDA tools with exceptional skill and efficiency to transform chip and system design,” said Cadence CEO Anirudh Devgan in the company’s press release.

A similar agentic theme runs through these announcements: the human sets intent and approves, the agent operates the toolchain, and the value is measured against rework, respin cost and closure time.

Data points worth pinning to a slide

• 3× — the stated speed advantage contributed by the optional machine-learning sketch filter, described as a factor of three.
• 4 parametric families plus boolean/sketch operations — the extrusion, revolution, fillet and chamfer set currently recoverable, alongside union, cut and sketch operations.
• 0 — plugins for common CAD software available today; a Siemens NX plug-in is in progress.
• 1 of 4 agentic announcements is positioned before schematic capture (Celus), while the others address execution, verification and tool driving.
• 4 vendors described by category: two EDA developers, one ECAD developer, one semiconductor design software developer.

Evaluation checklist before piloting any of these tools

• Classify your geometry population first: prismatic parts built from extrude, revolve, fillet and chamfer features fall inside today’s coverage, whereas loft-based and sweep-based bodies do not.
• Confirm licence posture in writing: open-source availability covers non-commercial use, and a patent application on the method is already on file.
• Test the interchange path, not just the demo: output lands in CadQuery Python, with Studio CAD as the companion viewer.
• Locate the human approval gate. For Celus the plan is executed only once the engineer approves it; document who that approver is in your own flow.
• Check what the agent is reading. Circuit Mind’s verification scope is datasheets, application notes plus your internal design review checklist — your checklist quality becomes a dependency.
• Ask which tools the agent can actually drive, and against which model set, using Cadence’s description of AI Super Agents operating EDA and SDA tools as the reference pattern.
• Keep commercial questions with the vendor: availability statements in the source material address capability and licence, not supply or contract terms.

Common misconceptions to correct in the design review

• “Mesh-to-CAD is a closed, proprietary capability.” CADFit sits in the same problem space as Backflip AI and Autodesk Fusion’s AutoTimeline, but is open-source for non-commercial use.
• “Editable history means full parametric parity with a human modelled part.” The stated operation set is bounded, and complex loft or sweep geometry is an acknowledged failure mode.
• “The machine-learning model is a required stage.” It is described as non-essential, contributing a factor of three in speed.
• “Agentic flows remove the engineer from the loop.” The ChipAgents framing keeps objectives, constraints and critical judgment calls with the engineer.
• “Verification assistants replace manual review.” Circuit Mind positions AVE as catching errors that slip past careful manual review, ahead of board respins — an additional gate, not a substitute.

FAQ

What does CADFit actually produce?

An editable construction sequence rather than a static solid: key sketches, converted solids, and the union and cut steps that consolidate them, expressed in CadQuery Python.

Where does it break down today?

On shapes outside its supported operations. Nehme’s position is that extrusions, revolutions, fillets, chamfers, union, cut and sketch operations are handled, while loft-based or sweep-based geometry remains difficult.

Can the agents ship a design without sign-off?

Not as described. Celus carries out the approved plan only after the engineer approves it, and ChipAgents positions the human as the owner of objectives, constraints and critical judgment calls.

What is Maestro, in practical terms?

A Cadence portfolio of domain-specific AI models for agentic chip and system design, which Anirudh Devgan says is combined with frontier and open-weight models so that AI Super Agents drive Cadence EDA and SDA tools.

Is “physics AI” the same thing as the agentic tools above?

No. In a guest article for Engineering.com, Luminary CEO Pete Schlampp separates physical AI from physics AI, and argues the second category has vast potential as AI that understands physics.

Action recommendations

1. Run a bounded CADFit trial on a representative slice of your mesh inventory — prismatic, feature-based bodies — and log the loft and sweep cases that fall outside current support, so the pilot result matches the documented capability.
2. Validate the CadQuery output path against your PLM and CAD environment before committing, given that no plugins for common CAD software exist yet.
3. Re-test the trial after the next CADFit version and the Siemens NX plug-in appear, since both are explicitly aimed at today’s gaps.
4. Pilot one agentic assistant at a time in the position each is documented to occupy — Celus ahead of schematic capture, ChipAgents across digital and analog execution, AVE at verification — with a named approver on the plan.
5. Feed AVE-class verification with your own design review checklist and current datasheet and application-note set, since that is the stated comparison basis.
6. Ask EDA suppliers for the same disclosure Cadence published for Maestro: which domain-specific models are used, which frontier or open-weight models complement them, and which tools the agents operate.

The remaining link: physical AI is not physics AI

Beyond tool releases, the distinction that will shape longer-term evaluation work is the one drawn in the Luminary guest piece hosted on Engineering.com: Pete Schlampp separates physical AI from physics AI, and frames the second as the more consequential category for engineering computation. Read against the CADFit paper and the agentic announcements, the practical lesson for semiconductor teams is consistent — capture intent-bearing geometry instead of baked surfaces, keep approval authority with the engineer, and treat each agent as a tool operator whose scope must be measured against your own rework and respin data.