cs.CESep 3, 2026
SaveLLM-Aided Design for Manufacturing: A Multi-Agent System for Intent-Preserving Redesign of CAD for Improved Manufacturability
Organizations: Department of Mechanical Engineering, Northwestern University, Evanston, USA · Department of Electrical and Computer Engineering, Northwestern University, Evanston, USA
Abstract
We introduce autonomous, intent-preserving Design for Manufacturing (DFM) redesign of CAD parts: given an engineer's CAD model, the method returns a variant that is easier to manufacture without losing its design intent. Generating such a redesign in a single shot is unreliable, since CAD fidelity degrades as parts grow complex; we instead produce it as a sequence of individually verified design transitions. Our DFM-Redesign pipeline realizes this with two coupled agent subsystems driven by a pretrained multimodal LLM: a DFM Reviewer that inspects the current design and proposes one intent-preserving manufacturability improvement at a time, and a CAD Modifier that executes each proposal as an edit to the part's CadQuery program. The CAD Modifier closes a verification loop, compiling every candidate edit and visually checking it against the intended change from multi-view renderings, then re-generating or re-instructing until the edit is accepted or abandoned. Iterating review and verified modification compounds edits into parts more complex than one-shot generators reliably produce, preserves the original intent at each step, and requires no fine-tuning. On a 46-part benchmark scored by chamfer distance to reference geometries, the CAD Modifier reproduces target parts more accurately on average than chain-of-thought single agents given the same tools, and ablations isolate the contributions of the visual review loop and of captioning the design state before each edit. A centrifugal pump casing built from 32 chained transitions illustrates the complexity reachable by compounding verified edits. This is a preliminary report: evaluation of the full review-and-redesign loop, including manufacturability gain and an operational measure of intent preservation, is ongoing.
Figures & tables
Figure 1: Intent-preserving DFM redesign of an example part. Starting from an initial design with raised (embossed) “A C” lettering (left), the DFM-Redesign pipeline applies a sequence of individually verified design transitions: Transition 1 converts the raised text to text engraved units into the surface, and Transition 2 adds -unit-radius fillets to the concave vertical corners of the characters. The final part (right) is more agreeable to CNC machining while retaining the original badge design intent.
Figure 2: The DFM-Redesign pipeline. A DFM Reviewer (a multimodal Gemini agent) inspects the current design—its CadQuery script, multi-view renderings, and textual part description—and emits a single manufacturability recommendation as a natural-language transition description. The downstream CAD Modifier executes that transition through three agents: an Initial Action Generator that produces a compile-verified “after” CadQuery script, a Visual Review Generator that checks the rendered result against the intended change, and a Re-Instructor that re-phrases the transition when visual validation repeatedly fails. The accepted “after” design becomes the input to the next review–modify cycle, compounding verified transitions until no further recommendation is warranted.
Figure 3: Step-by-step redesign of the badge part. Numbered panels trace the back-and-forth between the DFM Reviewer and the CAD Modifier. Transition 1 (embossed engraved text) is generated, compiler-verified, and visually accepted ( correct_execution: true ). Transition 2 (fillets on internal letter corners) succeeds on the first attempt. The third recommendation (chamfer the top edges of the base, ring, and disk) is repeatedly compiler-valid but rejected by the Visual Review Generator—because the inner edge of the ring is also chamfered—triggering the Re-Instructor. After the maximum number of retries the system abandons the transition and reverts to the last accepted state (Design State 3) as the final design.
Figure 4: Before, after, and delta for a slotted circular-boss part. Isometric views of the original design (left), the redesigned part after processing by the DFM-Redesign pipeline (center), and the highlighted difference (right, in red) between the two states. The redesign replaces hard-to-machine horizontal convex fillets with chamfers and adds fillets to the internal vertical slot corners; the corresponding transition instructions and rationales produced by the DFM Reviewer are enumerated in Table 1 .
| Transition 1 : | Transition 2: | Transition 3: | |
| title | Replace Horizontal Convex Fillets with Chamfers | Add Fillets to Internal Vertical Slot Corners | Replace Horizontal Fillets with Chamfers |
|---|---|---|---|
| description: | The horizontal convex edges located at the top of the three rectangular slots, which are oriented parallel to the Y-axis, are modified from 3.0-unit radius fillets to 45-degree chamfers with a leg length of 3.0 units. These chamfers define the transition between the top surface of the circular boss and the vertical walls of the slots. | The internal vertical corners of the three rectangular slots, which are 15 units wide and 125 units long, are updated to include fillets. Each of the four vertical edges in each of the three slots is modified with a fillet radius of 3.0 units. | The external horizontal fillets located on the top curved edges of the four segments of the central boss are replaced with 45-degree chamfers of 3.0 units. This change applies to the outer perimeter of the boss segments where the top face meets the cylindrical side wall, excluding the straight edges of the slots which are already chamfered. |
| rationale: | Horizontal convex fillets require specialized ball-nosed end mills and complex 3D tool paths to machine on a 3-axis mill. Replacing these fillets with 45-degree chamfers allows the part to be manufactured using standard chamfering tools and simpler 2.5D machining strategies, which significantly reduces machining time and production costs while still achieving the design intent of removing sharp edges. | These vertical corners are currently modeled as sharp 90-degree internal angles. Since machining is typically done with rotating cylindrical tools like end mills, a sharp internal vertical corner is impossible to achieve without expensive secondary operations like wire EDM. Adding fillets allows the slots to be machined efficiently with a standard end mill in a single setup, reducing manufacturing time and cost while also reducing stress concentrations in those corners. | Horizontal external convex fillets on curved paths require specialized ball-nosed end-mills and complex 3D toolpaths to machine the rounded profile. Replacing these fillets with 45-degree chamfers simplifies the manufacturing process by allowing the use of standard chamfering tools or basic milling operations with simpler 2D toolpaths, which significantly reduces machining time and cost while still eliminating sharp edges and reducing stress concentrations. |
Table 1: Transition instructions and rationales for the part in Figure 4 . Each column is one design transition produced by the DFM Reviewer, comprising a short title, the natural-language transition description passed to the CAD Modifier, and the manufacturability rationale generated alongside it.
Figure 5: Building a centrifugal pump casing over 32 chained transitions. The most complex part generated by the CAD Modifier to date. Each panel is one verified design transition applied to the state produced by the previous one; the reference geometry was downloaded from GrabCAD and the transition descriptions were produced by a manually-supervised LLM annotation pipeline.
Figure 6: Run-to-run variability under an identical transition sequence. (i) The reference STEP geometry, against (ii) an accurate and (iii) an inaccurate re-creation, each produced by running the same 32 transition descriptions. In run (iii) the visual review loop accepted an incorrectly oriented base cylinder, after which features described relative to that base—the taper ribbing among them—were mis-applied.
| raw | without anomalies | ||||||
|---|---|---|---|---|---|---|---|
| mean | median | std | anomalies | mean | median | std | |
| Variant A | 5.739 | 2.141 | 10.999 | [0, 30, 15, 24] | 2.943 | 1.700 | 3.673 |
| Variant B | 3.496 | 1.567 | 9.183 | [28] | 2.181 | 1.492 | 2.219 |
| Variant C | 4.474 | 2.047 | 9.482 | [28, 38] | 2.845 | 1.853 | 2.731 |
| Variant D | 11.495 | 2.578 | 29.746 | [27, 18, 1] | 4.087 | 2.492 | 5.819 |
| Variant E | 8.986 | 1.936 | 32.841 | [28, 0] | 3.329 | 1.775 | 4.286 |
Table 2: Ablation summary over the 46-part evaluation set. Chamfer distance statistics for the six CAD Modifier variants, reported over all parts (raw) and with the listed per-variant anomalies discarded. Best value in each column is bold; lower is better.
| Mean | Median | Std. dev. | Lower B | |||||
|---|---|---|---|---|---|---|---|---|
| A | B | A | B | A | B | Count | incid. | |
| Triggered | 6.229 | 2.411 | 3.576 | 1.386 | 7.552 | 2.758 | 12 | 8 |
| Non-triggered | 5.566 | 3.879 | 1.700 | 1.674 | 12.076 | 10.577 | 34 | 20 |
| Total | 46 | 28 | ||||||
Table 3: Re-Instructor ablation, split by whether the description revision was triggered. Chamfer distance statistics for Variant A (with periodic transition-description revision) against Variant B (without), separated into the 12 runs in which the revision fired and the 34 in which it did not. “Lower Variant B incidences” counts the parts on which Variant B achieved the smaller chamfer distance.
| Count | Avg. difference | |
|---|---|---|
| Lower Variant B incidences | 31 | 1.712 |
| Lower Variant C incidences | 14 | |
| Lower B by 1+ | 9 | 5.376 |
| Lower C by 1+ | 3 | |
| Lower B by 4+ | 6 | 7.017 |
| Lower C by 4+ | 1 |
Table 4: Effect of the pre-transition captioning step. Per-part win counts and mean chamfer distance differences between Variant B (with captioning) and Variant C (without), including the subsets where the margin exceeds 1 and 4 units.
Figure 7: Where the pre-transition caption helps. Parts on which Variant B (with the captioning step) tracks the reference geometry more closely than Variant C (without it).
Figure 8: Failure modes without the visual review loop. With no gating agent (Variant D), positioning and orientation errors go unchecked and compound across transitions, leaving final parts that diverge sharply from the design target.
| Count | Avg. difference | |
|---|---|---|
| Lower Variant B incidences | 29 | 8.939 |
| Lower Variant E incidences | 17 | |
| Lower B by 1+ | 9 | 28.453 |
| Lower E by 1+ | 2 | |
| Lower B by 4+ | 7 | 36.095 |
| Lower E by 4+ | 0 | 0 |
Table 5: Effect of supplying the history of previous transitions. Per-part win counts and mean chamfer distance differences between Variant B (no history) and Variant E (with an enumerated transition history).
| Variant B | Variant F | |
|---|---|---|
| Lower CD instances | 19 | 27 |
| Completed transitions | 219/221 | 220/221 |
| Failed reviews | 67/285 | 22/242 |
| Failed verifications | 87/373 | 38/281 |
| Average duration | 0:11:59 | 0:06:07 |
Table 6: Effect of the underlying Gemini model. Variant B (Gemini 3 Flash Preview) against Variant F (Gemini 3.5 Flash) over the 46-part evaluation set. Best value in each row is bold.
| raw | without anomalies | ||||||
| mean | median | std | anomalies | mean | median | std | |
| Variant B | 3.496 | 1.567 | 9.183 | [28] | 2.181 | 1.492 | 2.219 |
| Variant F | 3.634 | 1.398 | 10.030 | [28, 4] | 1.846 | 1.389 | 1.922 |
| Benchmark 1 | 6.812 | 1.384 | 22.135 | [27, 41] | 2.696 | 1.355 | 4.002 |
| Benchmark 2 | 41.992 | 2.170 | 194.168 | [40, 7, 27] | 4.926 | 2.072 | 7.492 |
| Benchmark 3 | 47.439 | 1.720 | 208.186 | [35, 20, 27] | 3.723 | 1.370 | 7.882 |
Table 7: Orchestrated CAD Modifier against chain-of-thought single-agent benchmarks. Chamfer distance statistics over the same 46-part evaluation set, with Variants B and F carried over from Table 2 for comparison. Best value in each column is bold; lower is better.
More in this topic
Multi-Agent LLM Systems
Multi-agent large language model systems coordinate multiple AI agents to solve tasks through communication, specialization, shared tools, and memory.
See all 1,001 papers