YOJERA
YOJ Engineering & Reasoning Assistant — from engineering data to engineering decisions.
A local, offline, privacy-first CAD and drawing ingestion, analysis and reasoning platform. Deterministic core plus three reasoning layers, with an opt-in local LLM. Working alpha — not a concept.
Why Design Change Analysis Is the Bottleneck
Millions of interconnected entities
A ship design holds millions of engineering entities bound together across structure, piping, electrical and outfitting.
One revision, many disciplines
A single design change propagates into systems that were never explicitly linked to it.
Analysis is manual and slow
Dependency tracing is done by hand, takes days, and leaves little traceable evidence behind.
Late detection is expensive
Impacts found downstream drive rework, production delays and avoidable project cost.
The Deterministic Core Is Running Today
Ingest, BOM, reconcile, diff and compliance — all deterministic, all offline, all shipping in the alpha.
- 2D ingest — PDF, TIFF, PNG, JPG, DXF, DWG
- 3D ingest — STEP, IGES, IFC (+ Inventor / SketchUp preview)
- BOM from 3D assembly, multi-level roll-up
- BOM from 2D drawing — table, grid and balloon
- Reconcile 3D BOM to drawing BOM, four-tier match
- Revision diff — envelope, solids, features, entity census
- Compliance engine — YAML rules, safe AST evaluation
- Change report — HTML, Markdown, PDF
- 2D SVG and 3D three.js viewers, part tree, isolate
- XCAF assembly, materials and PMI extraction
- SQLite catalog, self-migrating, ~34 columns
- Native-code isolation — forked child per CAD stage
reconcile and diff return exit code 1 when a model is out of sync or a revision has changed — so the platform drops straight into a CI or PLM release gate without a wrapper.
Two Hard Invariants
- Nothing leaves the machine.
- Geometry and numbers are never LLM-generated.
Three Reasoning Layers, Already Implemented
Cited rule retrieval, graph-based change impact, and a confidence-scored findings register.
Rule RAG knowledge base
- Structure-aware chunking of Markdown and PDF into document, section and clause units
- Hybrid retrieval — dense embeddings plus BM25, fused by reciprocal rank fusion
- Grounded generation: every answer resolves its citations back to real clauses
- Insufficient-evidence guardrail, plus a Recall@K evaluation script
Knowledge graph + change impact
- Typed, directed, evidence-carrying edges authored as YAML
- Deterministic breadth-first traversal from the changed object
- Returns affected entities, shortest dependency path, disciplines, downstream analyses and rules touched
- Offline interactive click-to-trace diagram
Findings, confidence & disposition
- Seven finding classes with severity and confidence tied to method
- Deterministic engines score 1.0; heuristic and LLM reads score lower and say so
- Accept / Reject / Modify / Defer / Refer, with note and full audit trail
- Re-running an analysis never overwrites a human decision
Demo content shipped with the alpha: ~100 clauses across 13 discipline documents, and a 43-node graph over four cross-linked subsystems — synthetic, for demonstration only.
AI Is Fenced In by Design
Input file
Router
by extension
Parser
Normalized artifact
SQLite catalog
+ JSON record
Analysis & reasoning
- Geometry, volumes, bounding boxes and counts
- BOM quantities and diff numbers
- Compliance verdicts and impact results
- Any number that appears in a report
One documented exception: the vision-BOM fallback may transcribe printed table values when OCR fails — flagged as model-read, low confidence.
- Temperature pinned to zero, every call schema-constrained
- No network calls, no telemetry, vendored JavaScript
- Local models only — served from localhost
- Geometry is never sent to the model
- All third-party components deliberately non-Chinese in origin
Local models: Gemma 3 12B · Phi-4-mini · Nomic-embed-text.
Two Interfaces Over One Engine
Command Line — cad-ingest
- ingest · bom · reconcile · diffDeterministic core; reconcile and diff exit 1 as gates
- rules build · ask · listCited answers from the local rule knowledge base
- impact analyze · graphTraversal results, or an offline interactive diagram
- findings gen · list · dispose · reportRegister, human disposition and dashboard export
Desktop GUI — Nine Tabs
The GUI is a thin veneer — every button calls the same module functions as the CLI, on background threads.
Where the Alpha Stands, Honestly
Version 0.0.1. The four stages run end to end; these are the limits we would close before any production use.
- Demo rule corpus and knowledge graph are synthetic — they must be replaced with signed-off engineering data
- Feature recognition (holes, fillets, chamfers) is heuristic and flagged as such
- An envelope-preserving change is invisible to the envelope layer; the entity census still moves
- No free local reader for SolidWorks, CATIA, NX or Creo — standardise on STEP AP242
- Findings adapters cover compliance, reconcile and impact; vision-BOM and RAG sources are follow-ons
- DWG needs the ODA converter on PATH; headless needs xvfb
A Working Alpha, Not a Concept Note
Engineering Benefits
- Faster design reviews
- Dependency analysis in minutes, not days
- Automated CAD to BOM validation
- Clause-cited compliance evidence
- Traceable, auditable engineering decisions
Organisational Benefits
- Lower engineering rework
- Reduced design cycle time
- Release gates that fail the build, not the ship
- Improved knowledge retention
- Data never leaves the yard network
Why YOJIX
Already built and runnable in v0.0.1:
- CAD and drawing ingestion
- 2D and 3D offline viewers
- BOM extraction and reconcile
- Revision diff and reporting
- Compliance rules engine
- Cited rule RAG assistant
- Knowledge graph and impact
- Findings and disposition