StudioPulse AI — Autonomous CAD Intelligence

Autonomous operations agent for CAD studios, turning live 3D telemetry into instant AI project quotes, real-time bottleneck detection, and automated client invoicing.

Comment

Inspiration: The $100B Custom Jewelry Blind Spot

In custom fine jewelry manufacturing, master artisans blend centuries-old craftsmanship with high-precision 3D CAD modeling in Rhino and MatrixGold. Sculpting bespoke cathedral shanks, organic pave galleries, and micro-claw stone seats requires hours of intense geometric modeling.

Yet operationally, custom jewelry CAD has always been an opaque, high-cost black box:

  • Zero Workshop Visibility: Studio managers have no real-time visibility into factory floor bottlenecks. When designers get stuck on non-manifold NURBS edge defects or repeating boolean subtraction failures, hours of productive time are silently lost.
  • Guesswork Quoting: Quoting custom client sketch photos takes hours of manual estimation, leading to under-priced labor or lost client orders.
  • Client Fee Disputes: High-end clients frequently challenge bespoke CAD design fees because 3D modeling looks deceptively simple to an untrained eye.
  • Catastrophic Crash Loss: Complex unmeshed SubD wireframes crash CAD workstations without warning, destroying hours of unsaved micro-surfacing.

We built StudioPulse AI to solve all four — creating an autonomous CAD telemetry intelligence and multi-agent platform powered by Google Cloud, Gemini 3.7 Flash Dynamic Thinking, and Gemma 3.

What It Does

StudioPulse AI connects directly to CAD workstations to capture real-time modeling telemetry, orchestrating an intelligent 4-agent ecosystem:

1. Autonomous Stuck Designer Rescue (4-Agent Mesh)

  • Gemma 3 Edge Telemetry Screener: Runs lightweight inference directly on the CAD workstation to screen high-velocity command streams (500+ commands/hour), filtering out routine drafting events to conserve cloud bandwidth and compute.
  • Gemini 3.7 Flash Dynamic Thinking: When an anomaly is detected (such as 3 boolean difference failures and 12 undos within 5 minutes), Gemini diagnoses the underlying topological defect (such as self-intersecting surfaces or naked edges) and synthesizes an actionable, 1-sentence Rhino coaching fix (such as ShowEdges or _ExtractSrf).
  • ADK Guardrail Approval Gate: High-severity interventions are held in a human-in-the-loop manager approval gate before being dispatched to the designer's screen.

2. Save-Guardian Proactive Crash Protection

  • Monitors active modeling velocity and elapsed time since the last explicit save.
  • When continuous complex SubD modeling exceeds 20 minutes (crash probability \(>80\%\)), Save-Guardian automatically triggers a disaster recovery snapshot to Google Cloud Storage (gs://studiopulse-cad-snapshots) and prompts an IncrementalSave.

3. Smart Pre-Fill AI Quoting & Empirical Calibration

  • Category-Aware Geometry Engine: The interface dynamically morphs across 6 jewelry categories — Rings (shanks and halos), Earrings (drop styles, lever backs, mirrored pairs), Pendants (bail types), Necklaces (chain styles), Bracelets (band styles), and Brooches (pin mechanisms).
  • Gemini 3.7 Flash Vision Extraction: Multimodal extraction identifies center stone cuts, total stone counts, and pave melee diamonds from 2D sketch photos.
  • Doubt Rating Engine: Single-angle visual ambiguities are flagged with amber Review badges and gemological reasoning. Managers verify with 1 click to confirm audited status.
  • BigQuery Empirical Grounding: Calculates labor hours grounded in 280+ historical project telemetry records using our empirical calibration formula:

T_estimated = T_base + Σ(N_i × t̄_i) + δ_studio

(Where T_base is setup time, N_i is stone count per type, t̄_i is BigQuery historical median setting seconds, and δ_studio is the studio-specific velocity factor).

4. Proof-of-Work Luxury Invoice Storyteller

  • Gemini Atelier Storyteller transforms raw CAD telemetry (\(24 \times \text{gvHalo}\), \(12 \times \text{BooleanDifference}\), \(2.33\text{ hours}\)) into an emotional Cartier-grade craftsmanship narrative on Page 1 of official printable client invoices — eliminating fee disputes by proving bespoke artisanal engineering.

5. Automated Excel (.xlsx) Reporting Engine

  • Generates Daily, Weekly, Monthly, and Custom Range audited spreadsheets for payroll and management with Smart Duplicate Reuse and a 30-day auto-expiry storage lifecycle.

How We Built It (Google Cloud Architecture)

[CAD Workstation] ---> [Gemma 3 Edge Screener] ---(Filtered Anomalies)---> [Cloud Run: FastAPI]
       ^                                                                           |
       |                                                                           v
(Dispatched Fix) <--- [ADK Guardrail Approval Gate] <----------------- [Gemini 3.7 Dynamic Thinking]
                                                                                   |
                                                           +-----------------------+-----------------------+
                                                           v                                               v
                                                [BigQuery Telemetry Lake]                       [Cloud Storage Snapshots]
                                                           |                                               |
                                                           +-----------------------+-----------------------+
                                                                                   v
                                                                      [React 18 + Firebase Hosting]
  • Compute: Containerized FastAPI backend deployed on Google Cloud Run in us-central1 with automatic scale-to-zero for zero idle cost.
  • AI Models:
    • Gemini 3.7 Flash (Dynamic Thinking Mode): Used for deep visual geometry extraction, CAD defect self-reflection, and luxury proof-of-work storytelling.
    • Gemma 3: Lightweight edge model deployed on workstations to screen high-velocity CAD command streams.
  • Agent Orchestration: Google Agent Development Kit (ADK) orchestrates the 4-Agent Mesh (Strategy Agent, Guardrail Agent, Save-Guardian, and Action Dispatcher) with human-in-the-loop tool confirmation gates.
  • Data and Storage:
    • Google BigQuery: studiopulse_lake dataset storing 280+ historical CAD project telemetry benchmarks.
    • Google Cloud Storage: studiopulse-cad-snapshots bucket for automated disaster recovery backups.
    • Cloud Firestore: Real-time synchronization of designer states and approval gates.
  • Frontend: React 18, Vite, Tailwind CSS, and Recharts, hosted on Firebase Hosting.

Challenges We Faced & Breakthroughs

1. Single-Angle 2D Sketch Ambiguity

  • Challenge: A single 2D sketch cannot show hidden pave stones beneath an under-gallery or the back of a piece, risking LLM hallucinations on pricing.
  • Breakthrough: We engineered the Doubt Rating Engine — Gemini assigns confidence levels (CONFIDENT, DOUBT, UNSURE) to every extracted parameter, highlighting ambiguous dimensions with amber Review badges and gemological explanations for instant 1-click verification.

2. Category Architecture Collisions

  • Challenge: Earrings, necklaces, and bracelets require completely different parameters than rings (drop styles and lever backs versus finger sizes and shanks).
  • Breakthrough: Built a dynamic CATEGORY_CONFIG state machine that adapts the UI layout, Gemini prompt engineering, and BigQuery labor calibration across all 6 jewelry categories from a single codebase.

3. Telemetry Signal-to-Noise Ratio

  • Challenge: A single designer generates 500+ CAD commands per hour. Streaming all raw events to Cloud Run would cause latency and cost spikes.
  • Breakthrough: Deployed Gemma 3 as an edge screener on the workstation, discarding routine commands locally and forwarding only genuine anomaly patterns to the cloud mesh.

Accomplishments We're Proud Of

  • End-to-End Working System: From the Rhino CAD tracker plugin to Gemma 3 edge screening, Cloud Run 4-agent mesh, and React 18 dashboard — every component is fully integrated and functional.
  • True Multi-Agent Orchestration: Not a single monolithic prompt, but four specialized agents collaborating through the Google ADK framework with human-in-the-loop safety gates.
  • 100% Automated Test Pass Rate: Verified through a 10-scenario evaluation suite (pytest tests/test_mesh.py) validating stuck rescue, quoting math, and guardrail routing.
  • Zero Idle Cost Architecture: Serverless Cloud Run and Firebase hosting ensure small jewelry studios pay $0 when workstations are closed.

What We Learned

  • Empirical Grounding Builds Trust: In manufacturing, generative AI is only trusted when grounded in real production data. BigQuery historical benchmarks transformed AI quotes from interesting guesses into verifiable estimates.
  • AI as a Coach, Not a Critic: Designers embrace actionable 1-sentence command suggestions (ShowEdges) rather than intimidating error logs.
  • Edge Screening is Essential: Gemma 3 pre-filtering on the workstation reduced cloud invocations by over 85%, ensuring real-time responsiveness.

What's Next for StudioPulse AI

  • Rhino 8 Grasshopper Parametric Sidecar: Real-time mesh defect healing directly inside the Grasshopper node canvas.
  • Multi-View 3D Mesh Reconstruction: Generating rotatable 3D preview meshes from 2D sketches prior to CAD drafting.
  • Enterprise Atelier Fleet Rollout: Centralized BigQuery analytics across multi-studio fine jewelry manufacturers.

Built With

Share this project:

Updates

Submission history