Skip to content

CAD Intelligence

Operational source:

C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\01_tracks\cad-intelligence

This page mirrors the detailed operational tracking for this track so the deployed wiki shows the same score tables as the source TRACKING.md.

Drawing-analysis system identity

Current model training, IONOS detector serving, the Railway validation fork, and the RapidDraft product are separate parts of the DraftLint System Family. Historical uses of “DraftLint” in this scorecard should not be read as one repository or runtime.

Last updated

2026-04-24

Scoring model

This track uses milestone and feature scorecards.

Score guide:

  • 100 means complete, validated, and stable for current scope
  • 90-99 means strong and working, but still carrying meaningful gaps
  • 75-89 means materially working, but still missing important completeness
  • 50-74 means real foundation exists, major parts are still absent
  • below 50 means early, partial, or mostly planned

For this mixed code-and-research track, 100 requires active plan evidence plus working implementation or evaluated benchmark results for the specific milestone.

Evidence sources reviewed

  • C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\01_tracks\cad-intelligence\MASTER_PLAN.md
  • D:\02_Code\45_merged_macos_colabui_dfmanim
  • D:\02_Code\49_yolotraining_firstdataset
  • D:\02_Code\50_CVAT_RoboFlow
  • C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\01_tracks\cad-intelligence\handover\260315_design-review-mode-boundary.md
  • C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\01_tracks\cad-intelligence\handover\260318_candidate-viewer-guardrails.md
  • C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\01_tracks\cad-intelligence\handover\260407_part-facts-refresh-latency-plan.md
  • C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\01_tracks\cad-intelligence\handover\260408_rule-driven-review-preparation-pipeline-plan.md
  • C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\01_tracks\cad-intelligence\handover\260408_cad-intel-perf-accuracy-plan.md
  • C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\01_tracks\cad-intelligence\handover\260408_dfm-review-pdf-export-plan.md
  • C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\01_tracks\cad-intelligence\handover\260408_railway-one-service-simplification.md
  • C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\01_tracks\cad-intelligence\handover\260417_cnc-physics-family-rule-rollout-plan.md
  • C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\01_tracks\cad-intelligence\handover\260418_dfm-rules-and-localization-guidance.md
  • C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\01_tracks\cad-intelligence\handover\260424_dfm-scanner-pipeline-dev-loop.md
  • D:\02_Code\45_merged_macos_colabui_dfmanim\docs\validation\injection-molding-dev-loop-review.md
  • D:\02_Code\45_merged_macos_colabui_dfmanim\docs\contracts\dfm-scan-pipeline.md
  • D:\02_Code\45_merged_macos_colabui_dfmanim\docs\validation\dfm-scan-benchmark.md
  • C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\04_playbooks\DRAWING_MODEL_TRAINING_AND_CVAT_MODEL_SERVING.md
  • C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\TextCAD_Wiki\docs\04_DFM_Research\_INDEX.md

Current overall score

CAD Intelligence overall score: 80 / 100

Why the score is not higher: The integrated runtime is now much more real and pilot-usable than it was in March. The largest new gain is injection molding: the plastics route now has a rules-bar process view, Light/Deep scan control, scan-depth-aware Part Facts, product-vs-tooling gating, richer boss/rib/draft/side-action evidence, and a validated six-sample STEP corpus. The April 24 scanner-pipeline loop also made heavy STEP review more transparent with scanner-level progress, cache state, selected-component scope checks, and a first bounds-aware multipart component resolver. The labeling and training operations around CVAT, Vast.ai, W&B, YOLO, and Roboflow are also documented and materially working enough to support repeatable experiments. The score is not higher because deep scan can still be slow on heavy geometry, Part Facts is still too much of an orchestrator, topology-linked localization is not solved, and bounded geometry edits remain absent.

Milestone scorecard

Milestone Score / 100 Current state What it helps achieve Main remaining gap
1. DFM benchmark foundation 88 Strong and expanding Provides the core manufacturability intelligence and rule discipline. Still needs tighter benchmark reporting alignment across CNC and plastics paths.
2. Feature detection and classification 79 Materially working Identifies manufacturability-relevant evidence families for active review. Accuracy and confidence still vary on harder geometry and assemblies; long OCC detectors still need scanner-unit extraction.
3. Geometry localization 52 Improving but not solved Links findings back to real geometry for review and later remediation. Selected-component scope is safer, but topology-linked targeting is still not trustworthy enough.
4. Injection-molding benchmark 74 Materially working Extends the track beyond CNC with an active plastics rules bar and Light/Deep scan validation path. Still needs broader scored benchmark discipline and hosted runtime verification.
5. Bounded edit loop 18 Very early Enables one trusted AI-assisted geometry edit path. No production-worthy direct edit loop exists yet.
6. Labeling and evaluation operations 74 Materially working Supports repeatable data creation, training, and CVAT-served detector experiments. Model quality and dataset scale still lag the now-stable operating workflow.

Feature scorecard

Feature track Score / 100 What it helps achieve Current note
CNC DFM 86 Gives the product a strong manufacturability baseline. Still the deepest and most trusted capability family.
Injection-molding DFM 74 Broadens manufacturing coverage. Now has 12 active plastics rules, five J-book plastics rules, Light/Deep scan, tooling gate, and six-sample validation.
Feature detection 79 Helps turn geometry into actionable engineering facts. Deep scan now extracts richer molded bosses, ribs, draft, side actions, and sink/mass screening evidence; scanner-level timings now expose where heavy work is happening.
Geometry mapping 52 Makes findings spatially trustworthy. Plastics validation localizes rule violations well, and multipart selected-component scans now have bounds/scope guards, but topology-linked remediation targeting is still not solved.
Labeling pipeline 74 Supports repeatable data creation. CVAT, Vast.ai, W&B, and model-serving paths are now documented and reusable, but data coverage is still thin.
Benchmark reporting 78 Makes progress measurable across runs. Injection molding now has a concrete harness, but cross-process scoring/reporting still need consolidation.
Bounded edit capability 18 Enables AI-assisted geometry modification. Still more aspiration than validated workflow.

Current headline assessment

CAD Intelligence is no longer just a CNC-heavy benchmark foundation. It is now a real integrated review runtime with rule-driven preparation, selective evidence planning, persisted Part Facts warmup, Light/Deep scan control for injection molding, plastics product-vs-tooling gating, scanner-level progress, selected-component scope validation, pilot-facing outputs such as DFM PDF export, and a documented drawing-model operations stack that spans CVAT labeling, Vast.ai training, W&B run tracking, and CVAT-linked model testing.

The track is still held back by geometry-localization trust and the absence of one bounded edit workflow that can be defended end to end.

Active rollout updates

2026-04-24 - DFM scanner-pipeline modernization loop

The DFM review pipeline is being moved from evidence-family-only progress toward scanner-level transparency and safer selected-component behavior.

Current rollout policy:

  • keep review-v2 and Part Facts payload shapes stable while adding scanner status fields
  • expose planned scanners, active scanner, elapsed time, budget/overtime state, timing, and cache state
  • keep Light scan as the default fast path and use GE_JET_BRACKET_V_3.0.stp as the heavy regression fixture
  • make multipart STEP review selected-component scoped by default
  • reject escaped scanner evidence before it becomes misleading issue cards
  • refactor Part Facts into scanner artifacts only after additive status/scope behavior is validated

Completed April 24 slices:

  • scanner status rows now carry started_at, budget_ms, overtime_ms, and cache state
  • scanner planning has table-driven coverage for CNC light/deep, plastics, sheet metal, drawing-only, and assembly-only routes
  • review orchestration has started moving out of main.py into a dedicated pipeline class
  • CNC corner extraction can reuse the already loaded resolved OCC shape instead of reloading the STEP
  • scanner artifacts are written atomically and expensive wall/molded-signal outputs have cache artifacts
  • selected-component scope metadata prevents multipart escaped evidence from being presented as valid review findings
  • single-component GE scans are exempt from multipart-only scope rejection
  • selected-component resolution now prefers OCC solids whose CAD bounds match the selected scene component before falling back to ordinal component_N

Validation:

  • focused CNC/Part Facts bounds resolver tests: 56 passed
  • broader DFM/canonical scene subset after bounds resolver: 79 passed, 1 skipped
  • previous full touched DFM backend suite: 140 passed
  • frontend build in web: passed with the existing chunk-size warning
  • GE heavy benchmark through local API: Light review 269.4s, Deep review 467.7s, repeat cached Deep review 3.1s

Remaining risk:

The scope guard and bounds-aware resolver improve safety, but Deep scan is still too slow on heavy geometry. The next architecture loops are scanner-unit extraction, Part Facts as a materialized view over scanner artifacts, and progressive/deferred Deep scan.

2026-04-21 - Injection molding Light/Deep scan and plastics rule coverage active

Injection molding now has a credible active review path in the RapidDraft integration repo.

Current rollout policy:

  • plastics and injection_molding resolve to the same Injection Molding rules-bar view
  • 12 plastics rules are active by default: 6 implemented and 6 heuristic screening rules
  • J_BOOK_PLASTICS contributes five boss/rib rules to the plastics route
  • Light scan remains the default fast review path
  • Deep scan adds richer ribs, bosses, collars, pins, gussets, draft, side-action grouping, sink/mass screening, bbox localization, and face-index localization
  • mold/tooling assemblies are gated out of product-part findings and receive a clear selection warning
  • validation harness runs both Light and Deep scan over the six STEP samples, with Deep scan completing 6/6 and localizing all rule violations in the validation summary

2026-04-13 - Drawing-model training and CVAT serving workflow captured

The drawing-analysis side of the track now has a reusable operating path instead of scattered local memory.

Current rollout policy:

  • CVAT on the Fedora mini PC is the shared labeling and model-serving surface
  • Nuclio-backed model functions must preserve the trusted-origin setup for both cvat_server and cvat_worker_annotation
  • Vast.ai is the preferred GPU training path for YOLO experiments
  • W&B is the preferred remote run-history and dashboard surface
  • YOLO and Roboflow experiments must be isolated by branch or worktree, but they share the same CVAT runtime constraints

Highest-value next actions

  1. Extract the heavy scanner units (face_inventory_exact, cnc_corner, internal_radii) into independently cacheable and cancellable modules.
  2. Convert Part Facts from geometry orchestrator into a materialized view over scanner artifacts.
  3. Make Deep scan progressive/deferred so Light review returns quickly and Deep evidence enriches transparently.
  4. Replace box-only blame mapping with topology-linked localization.
  5. Define the first editable feature class and the acceptance criteria for its edit loop.

Sources

  • C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\01_tracks\cad-intelligence\MASTER_PLAN.md
  • C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\01_tracks\cad-intelligence\TRACKING.md
  • D:\02_Code\45_merged_macos_colabui_dfmanim\docs\contracts\dfm-scan-pipeline.md