Chamfer AI — Pitch Deck Slide 1 — Title Chamfer AI Describe a part in English. Our agent writes the CAD code, runs it, inspects the geometry, and repairs itself until the part is right. Sub-line for the opening beat: AI transformed software, art and finance. Mechanical engineering — where every physical product begins as a CAD model — has not yet had that transformation. That's the gap we're building into. Slide 2 — Problem Title as claim: General-purpose LLMs are bad at CAD — and they fail in three specific, expensive ways. 1. They guess dimensions. "Make me a mounting bracket for a motor" is how a real engineer actually talks. That prompt is under-specified. A general LLM doesn't ask — it invents a hole diameter and moves on. The ProCAD paper measured this directly: on a 2,469-sample ambiguous-prompt test set, Claude Sonnet 4.5 as a single model produced an invalidity ratio of 14.6%. The failure mode is under-specification, and the model cannot see it. 2. The code is often invalid or unbuildable. CAD scripting needs exact selectors, valid solids, legal operation order. General models hallucinate all of it. Independent benchmarks agree: the MUSE benchmark found LLMs "fail to reliably satisfy functional, manufacturing, and assembly requirements"; Text2CAD-Bench says text-to-CAD "remains largely unsolved beyond basic geometry," with chamfers, fillets, sweeps and lofts largely absent. A Cambridge study had GPT-4o produce correct geometry on only 10 of 15 bracket tasks. 3. It's expensive and slow. Running geometry synthesis on a frontier model has real latency and real token cost. Zero-to-CAD's own base-model result shows the gap starkly: off-the-shelf Qwen3-VL-2B scored a 6.6% success rate before fine-tuning. The cost of brute-forcing that with a frontier model is what kills interactive design. And the stakes are physical. A mesh or a guess is not acceptable here. Engineering needs parametric, editable, manufacturable geometry. A design error caught at requirements costs 1 × 1×; caught in production, up to 1500 × 1500× (NASA/Stecklein, peer-reviewed). 86% of rework cost is locked in before design freeze (Tan, Otto & Wood 2017). Bad geometry doesn't fail on a screen — it fails on a test rig four months later. Slide 3 — The technical thesis: fine-tuned CAD models, not bigger general models Title as claim: Two 2026 papers show the same thing — a domain fine-tuned model beats a frontier model at CAD, at a fraction of the size. Left chart — code execution success rate. Zero-to-CAD (Autodesk Research, arXiv 2604.24479, Apr 2026) reports on its own test set: Model Success rate Mean IoU Qwen3-VL-2B, fine-tuned 82.1% 0.747 GPT-5.2 (High) 72.2% 0.485 Qwen3-VL-2B, base 6.6% 0.184 A 2-billion-parameter model, fine-tuned on CAD data, beats GPT-5.2 by 9.9 points on success and roughly doubles it on geometric fidelity (IoU 0.747 vs 0.485). The base model of the same architecture scores 6.6% — so essentially all of that capability is the fine-tune, not
Chamfer AI — Pitch Deck Slide 1 — Title Chamfer AI Describe a part in English. Our agent writes the CAD code, runs it, inspects the geometry, and repairs itself until the part is right. Sub-line for the opening beat: AI transformed software, art and finance. Mechanical engineering — where every physical product begins as a CAD model — has not yet had that transformation. That's the gap we're building into. Slide 2 — Problem Title as claim: General-purpose LLMs are bad at CAD — and they fail in three specific, expensive ways. 1. They guess dimensions. "Make me a mounting bracket for a motor" is how a real engineer actually talks. That prompt is under-specified. A general LLM doesn't ask — it invents a hole diameter and moves on. The ProCAD paper measured this directly: on a 2,469-sample ambiguous-prompt test set, Claude Sonnet 4.5 as a single model produced an invalidity ratio of 14.6%. The failure mode is under-specification, and the model cannot see it. 2. The code is often invalid or unbuildable. CAD scripting needs exact selectors, valid solids, legal operation order. General models hallucinate all of it. Independent benchmarks agree: the MUSE benchmark found LLMs "fail to reliably satisfy functional, manufacturing, and assembly requirements"; Text2CAD-Bench says text-to-CAD "remains largely unsolved beyond basic geometry," with chamfers, fillets, sweeps and lofts largely absent. A Cambridge study had GPT-4o produce correct geometry on only 10 of 15 bracket tasks. 3. It's expensive and slow. Running geometry synthesis on a frontier model has real latency and real token cost. Zero-to-CAD's own base-model result shows the gap starkly: off-the-shelf Qwen3-VL-2B scored a 6.6% success rate before fine-tuning. The cost of brute-forcing that with a frontier model is what kills interactive design. And the stakes are physical. A mesh or a guess is not acceptable here. Engineering needs parametric, editable, manufacturable geometry. A design error caught at requirements costs 1 × 1×; caught in production, up to 1500 × 1500× (NASA/Stecklein, peer-reviewed). 86% of rework cost is locked in before design freeze (Tan, Otto & Wood 2017). Bad geometry doesn't fail on a screen — it fails on a test rig four months later. Slide 3 — The technical thesis: fine-tuned CAD models, not bigger general models Title as claim: Two 2026 papers show the same thing — a domain fine-tuned model beats a frontier model at CAD, at a fraction of the size. Left chart — code execution success rate. Zero-to-CAD (Autodesk Research, arXiv 2604.24479, Apr 2026) reports on its own test set: Model Success rate Mean IoU Qwen3-VL-2B, fine-tuned 82.1% 0.747 GPT-5.2 (High) 72.2% 0.485 Qwen3-VL-2B, base 6.6% 0.184 A 2-billion-parameter model, fine-tuned on CAD data, beats GPT-5.2 by 9.9 points on success and roughly doubles it on geometric fidelity (IoU 0.747 vs 0.485). The base model of the same architecture scores 6.6% — so essentially all of that capability is the fine-tune, not
Created using ChatSlide
Chamfer AI bridges the gap in CAD transformation by enabling users to describe parts in English, allowing AI to generate precise CAD code. It enhances design processes by inspecting and self-repairing geometry, turning AI advancements into tangible designs. However, general LLMs often falter in engineering due to ambiguous prompts leading to costly errors and unbuildable geometry. In contrast, fine-tuned CAD models significantly outperform, achieving 82.1% accuracy compared to GPT-5.2's...