FabVerdict: a manufacturability verdict and fair-price band on every CAD file, before money changes hands

The problem

DFM feedback is the most expensive missing step in on-demand manufacturing. A bad file means a rejected quote, a wasted tooling charge, or a part that arrives wrong weeks later. Beginners burn most of a filament spool on failed prints, r/3DPrinting is a museum of failures, and CNC shops spell out how much drawing review must happen before a quote is even valid. The feedback exists — but only inside giants' walls.

The idea

Why now

DFM feedback used to need a human applications engineer (Xometry's model — expensive, platform-locked) or heavyweight geometry kernels. The text-to-cad agent skill library (16.6k stars) gives any open agent the eyes: plain-language CAD inspection, validated STEP/STL, slicing, DfAM checks. The new local decision models (ARC-1 at sub-25ms, this week's 41M-parameter TokenAI Neo) give it the judgment at ~$0 marginal cost per file. The combination turns 'upload, wait for a human' into 'upload, verdict in seconds' for any fab, not just the giants.

What it combines

Two radar capabilities combine: (1) open CAD agent skills (text-to-cad: CAD, CAE, CAM) as the geometry-inspection engine that can look at, slice, and DfAM-check any file; (2) local typed decision models (ARC-1, JEV-27B) as the verdict layer that scores pass/fail plus severity cheaply enough to run on every upload. The mix matters because inspection without judgment is a raw checklist nobody acts on, and judgment without inspection is a guess — together they are a verdict a shop can price against.

MVP

Build: drag-and-drop STL upload, text-to-cad DfAM checks, and an ARC-1 verdict card (pass, warnings list, fail with reasons) with estimated print-time and material band. Skip: CNC and sheet-metal processes (FDM/SLA only), the fair-price band model (show material plus time cost breakdown instead), STEP import, and the marketplace API.

Distribution

B2B2C via fab marketplaces and CAD plugins: the long tail of independent fabs, regional job shops, and niche marketplaces pay per verdict or flat monthly; second wedge is procurement teams who want a second opinion on supplier quotes.

Why it wins

Xometry's DFM is a moat for Xometry; FabVerdict is Switzerland — an embeddable verdict any fab marketplace or CAD plugin can ship in an afternoon, with the fair-price band as the trust anchor (shops cannot be accused of padding quotes when the verdict is third-party).

Risks

DFM judgment quality — a wrong 'pass' that leads to a failed $2,000 CNC order destroys trust instantly. De-risked by scoping the MVP to FDM/SLA (where the checks are strongest and failure cost is low) and defaulting to conservative verdicts with explicit 'uncertain — human review' states.

Build it with

• text-to-cad: agent skills for CAD, CAE and CAMThe open agent skill library for CAD inspection, slicing, and DfAM checks — the geometry eyes.
• ARC-1: 1.7B local decision model for typed choices, scores and yes/no judgmentsSub-25ms local typed judgments (pass/fail plus severity) cheap enough to run on every upload.
• JEV-27B: open System 1 decision layer grafted onto Qwen3.8-27BCalibrated confidence for the verdict card's explicit 'uncertain — human review' states.

Repo to start from

fabverdict-checks — open DfAM check definitions (overhang angles, wall thickness, tolerance flags per process) plus the verdict-card schema; the calibrated scoring model and price-band data stay closed.

Evidence

• Formlabs launches biggest print disaster contest on Reddit (VoxelMatters)
• HN: 3D printing reliability for beginners
• Xometry Instant Quoting Engine