Drafty Autopilot: Engineering-Aware CI/CD

Inspiration

AI coding agents can generate software faster than ever. But there's a critical problem: code that builds successfully doesn't necessarily produce correct engineering results.

In CAD applications, a small software regression can silently change dimensions, break geometry, or produce invalid 3D models, even when conventional CI/CD tests pass.

We built Drafty Autopilot to bridge the gap between AI-assisted software development and verifiable engineering correctness.

What it does

Drafty Autopilot combines GitLab Duo Agent Platform, deterministic CAD validation, and supervised DevSecOps to detect and repair engineering-specific software regressions.

Our solution introduces three core capabilities:

1. CAD-Aware Validation

OpenSCAD and Trimesh independently validate generated 3D geometry, checking mesh integrity, watertightness, volume, and dimensional tolerances against available engineering specifications.

2. Evidence-Driven Self-Healing

When a regression is detected, GitLab CI produces structured JSON reports containing measured values, expected dimensions, and failed checks.

A custom GitLab Duo agent analyzes the evidence, identifies the root cause, proposes a targeted correction, and creates a Merge Request for independent verification.

3. Supervised Software Delivery

GitLab CI/CD combines software testing, CAD validation, security scanning, Docker packaging, and controlled deployment workflows with human approval checkpoints.

Our real GitLab Duo demonstration

We demonstrated self-healing using an intentionally introduced geometric regression.

In MR !5, a perforated plate fixture was changed from its required width of 100 mm to 90 mm.

The frontend build and backend tests passed, but CAD-Aware Validation detected the dimensional mismatch and rejected the geometry.

Our GitLab Duo Self-Healing Agent analyzed the failed validation report, identified the incorrect parameter, and generated MR !6 to restore the correct dimensions.

The repair pipeline passed, confirming the corrected geometry while retaining the negative regression test.

This is a real agent-executed repair, supported by GitLab session logs, Merge Requests, and CI evidence.

How we built it

We selected Path B: Bring Your Own Project, extending the existing MIT-licensed Drafty application.

Our updated architecture uses:

React, Vite, Three.js, and FastAPI for the application.
OpenAI for image understanding and parametric OpenSCAD generation.
OpenSCAD and Trimesh for deterministic geometric validation.
GitLab Duo Agent Platform for supervised AI-driven diagnosis and code remediation.
GitLab CI/CD for builds, tests, security analysis, and software delivery.
Docker for reproducible container packaging.
Google Cloud Run for application hosting, with Google Cloud Storage support for persistence.

Our updated GitLab pipeline has passed 37 automated tests, with no failures and two skipped tests, and has successfully built and published a Docker image.

Challenges

The biggest challenge was connecting probabilistic AI reasoning with deterministic engineering requirements.

A language model may propose a plausible correction, but correctness must be established independently.

We addressed this by separating AI-generated changes from OpenSCAD and Trimesh measurements and requiring repeatable validation evidence.

We also modernized the original application by replacing legacy Huawei ModelArts and MongoDB dependencies with a simpler OpenAI-based architecture.

What we learned

AI agents become significantly more useful when their decisions are connected to measurable, reproducible evidence.

For engineering software, passing conventional tests is only part of correctness. Validating what the software actually produces is equally important.

What's next

We plan to expand CAD-Aware Validation to additional parametric models, support richer engineering specifications, improve production monitoring, and extend evidence-driven self-healing to other engineering domains.