What we built, what 1,000 files proved, and why we paused.
A transparent engineering report on Project La Vinci: our DWG ingestion stack, automated execution versus manual source fidelity, preserved assets, and Rine's next learning direction.
Decision recorded 10 October 2026 · Current parser build path formally paused
Sunk engineering effort is not product strategy.
La Vinci began as an ambitious attempt to create a universal, reusable engineering-file layer around a normalized intermediate representation. We engineered a headless DWG ingestion pipeline, downstream vector conversion, and high-resolution rendering tools.
Across weeks of testing against a dedicated 1,000-file corpus, the pipeline proved that complex CAD files could be processed deterministically in cloud environments without desktop runtime dependencies.
However, rigorous manual inspection revealed a critical gap: successful execution is not the same thing as source fidelity. Furthermore, repeated customer conversations demonstrated that another general-purpose CAD conversion engine is not the primary infrastructure problem software companies need solved.
What Project La Vinci Implemented
Headless DWG/DXF Ingestion
Low-level extraction parser reading binary DWG, ASCII DXF, and template DWT files without AutoCAD or external runtime licenses.
LAVINCI_CAD_IR_V3
A normalized Intermediate Representation schema translating proprietary CAD entity trees, block dictionaries, and layer hierarchies into standard JSON.
Multi-Target Output
Downstream generation pipelines rendering scale-accurate vector PDF, clean web SVG, normalized exchange DXF, and high-resolution raster tiles.
The 1,000-File Validation Test: What It Proved vs. What It Did Not
What automated testing proved
- 01Execution resilience: Complex industrial DWG files can be ingested in serverless containers without crashing or memory exhaustion.
- 02Normalization viability: Geometry entities, text entities, and block definitions can be mapped into a unified intermediate schema.
- 03Diagnostic observability: Parsing warnings, unsupported entity notices, and coordinate bounds can be exposed programmatically.
What it did NOT prove (The Fidelity Failure)
- 01Execution ≠ Source Fidelity: A job completing with exit status 0 does not mean the rendered output matches the author's exact CAD viewport.
- 02Visual inspection discrepancies: Manual comparison uncovered dimension line misalignments, hatch style variances, and complex linetype discrepancies.
- 03Automated metric blindness: Automated regression suites evaluated schema integrity, not human perceptual fidelity or engineering intent.
Preserved, not discarded.
Pausing active API expansion does not mean discarding engineering work. The La Vinci assets provide Rine with durable technical capability as we refocus on real customer workflows.
The 1,000-File Validation Corpus
A curated benchmark set of real-world CAD drawings spanning architectural plans, mechanical parts, and electrical layouts.
LAVINCI_CAD_IR_V3 Schema
Formal TypeScript and JSON Schema definitions for representing CAD drawing primitives, blocks, and entity graphs.
Parser & Conversion Pipeline Codebase
Complete parser implementations, vector rendering routines, and diagnostic collectors preserved in cold storage.
Automated Regression Test Suite
End-to-end testing harness evaluating memory bounds, throughput metrics, and entity counts.
Where Rine is investigating next.
The strongest customer pull is pointing toward a distinct boundary: engineering information inside drawings, specifications, attachments, and technical documents needing to become structured data inside another software workflow.
Software teams usually already have a way to store files. What they lack is an accurate, reliable way to extract technical information and feed it into estimation pipelines, verification tools, or AI workflows. Further product development will earn its way into the roadmap through customer evidence.