The Autonomous, Grounded Hardware Diagnostic Engine.
A multi-modal infrastructure that teaches you to build for the physical world.
🌐 Live Demo: prox.ashishsandhu.com
[High-fidelity hands-free diagnostic flow]
The Vulcan OmniPro 220 is not a simple tool. It supports MIG, flux-core, TIG, and stick. It runs on 120V and 240V. Each process has different polarity, gas, wire-feed, and duty-cycle rules.
The manual has the answers, but the answers are spread across tables, setup diagrams, troubleshooting charts, process-selection pages, and weld defect images.
Vulcan-OS (VO) turns that manual into a garage-side support experience. It doesn't just answer questions; it directs the user through high-risk hardware configurations with the precision of a senior welding engineer.
VO uses a Modular Multi-Agent Ecosystem. We separate Direction (AI Reasoning) from Execution (Deterministic Rendering).
graph TD
User([User Input + Image]) --> Planner{Tier 0: AI Planner}
Planner -->|Sub-100ms Heuristics| Heuristics[Groq / Gemini Flash]
Planner -->|Grounded Intent| Controller[Director Engine]
Controller -->|Fetch Facts| Registry[(Hardware Fault Registry)]
Controller -->|Fetch Manual| Manual[(Manual Visual Registry)]
Registry --> Reasoning{Tier 1: Reasoning}
Manual --> Reasoning
Reasoning -->|Deep Logic| Claude[Claude 4.6 Opus / Gemini 3.1 Pro]
Claude -->|Visual Spec| Workspace[Diagnostic Workspace]
Workspace -->|Render| Topology[Digital Twin Topology]
Workspace -->|Render| Calculator[Duty Cycle Matrix]
Workspace -->|Render| Troubleshooting[Diagnostic Checklist]
subgraph "The Truth Layer"
Claude -.->|MCP Tool Call| Server[Custom MCP Server]
Server --> Registry
end
Every AI model hallucinations. In a welding garage, a hallucinated polarity setup can destroy a machine or ruin a workpiece. Vulcan-OS solves this with a Safety Arbiter:
- Truth Over Prose: If the LLM generates a response that contradicts the hardcoded safety facts in
ProductGrounding.ts, the system clobbers the AI response and replaces it with the canonical, verified setup recipe. - Pre-emptive Grounding: Critical facts (like duty-cycle limits) are auto-prepended to answers by the application layer, ensuring the user sees the "Truth" before the "Narrative."
Vulcan-OS is a product-agnostic orchestration layer. While currently grounded in the Vulcan OmniPro 220, the system is designed as a "compiler" for any technical hardware product.
Transitioning VO to a new domain (e.g., a 3D printer, CNC mill, or medical device) requires zero architectural changes—only a data injection:
- Hardware Fault Registry: Swap the 15+ welding fault nodes in
src/data/for the new product's error codes (e.g., "Extruder Clog", "Z-Axis Limit"). - Product Grounding: Update the technical constraints (thermal limits, power specs, material matrices) in
src/data/ProductGrounding.ts. - Visual Registry: Map new manual diagrams and component photos in
src/data/VisualRegistry.ts. - Deterministic Engines: Update the "Quick Setup" decision tree in
src/engines/quickSetup.tsto reflect the new product's initialization sequence.
Infrastructure-First Outcome: The "Director" logic (streaming, voice, multimodal intent routing) remains untouched. You aren't building a single-use app; you're deploying a Universal Technical Support OS.
Every hardware assistant on the market makes a fundamental error: They treat documentation as a text-retrieval task.
- Red-Teaming Constraints: We treat hardware specs (Polarity, Duty Cycle, Gas Flow) as Hard Constraints. The AI is not allowed to "suggest" these; it must enforce them.
- Deterministic Grounding: By separating the "Brain" (LLM) from the "Truth" (Manual Registry), we ensure that the system is 100% accurate on safety-critical data.
- No Hallucinations: We use a Digital Twin Topology Engine. The AI explains the logic, but the Application Layer controls the physical coordinates of the wiring diagram.
The Problem: AI models are notoriously bad at spatial reasoning. If you ask an LLM to "draw a wiring diagram," it will likely hallucinate cable paths. The Solution: VO uses a Digital Twin. We have mapped every physical port of the Vulcan OmniPro 220 to a pixel-perfect React coordinate system.
sequenceDiagram
participant User
participant AI as AI Director
participant Twin as Digital Twin Engine
User->>AI: "How do I wire for TIG?"
AI->>AI: Identify Process (TIG)
AI->>Twin: Request Topology(process="tig")
Twin-->>Twin: Lookup X/Y Socket Coordinates
Twin-->>AI: Validated Terminal Map
AI->>User: "Connect Torch to Neg, Ground to Pos." (Rendered Visual)
Vulcan-OS uses a layered reliability model to ensure garage-side stability:
- Custom MCP Server: Built on the Claude Agent SDK, our
vulcanMCP server exposes programmatic tools (lookup_vulcan_manual_context) that allow Claude to dynamically fetch manual evidence with sub-page precision. - Hybrid Voice Pipeline: Optimized for "Gloves-On" utility.
- Immersive Mode: A continuous Whisper-based capture loop for power users.
- Wake-Word Mode: Browser-native listener for "Hey Vulcan" that triggers high-fidelity transcription.
- Topology Coordinate Mapping: The Digital Twin is not a generated SVG; it is a React engine where every physical port is mapped to pixel-perfect X/Y coordinates on the product image.
 [Digital Twin Topology Engine] |
 [Multi-Modal Vision Audit] |
- Answer polarity questions and draw the cable setup deterministically.
- Show the real manual image when the answer depends on a specific diagram.
- Calculate duty cycle and convert it into weld/rest time.
- Compare MIG, flux-core, TIG, and stick setups in a high-fidelity table.
- Render pre-weld checklists for one process or all four processes.
- Walk through porosity, spatter, wire-feed, and bad-weld troubleshooting.
- Diagnose uploaded weld photos with Claude vision.
- Recommend a process from location, gas, material, and thickness.
- Surface fault states like HOT, low voltage, over voltage, no power, no arc.
- Deterministic Response Normalization: Automatic clobbering of LLM hallucinations if they contradict safety registries.
The manual is represented as structured product knowledge instead of raw PDF chunks.
src/data/HardwareFaultRegistry.ts- 15+ Proprietary failure nodes.src/data/ProductGrounding.ts- Polarity, duty cycle, and setup facts.src/data/VisualRegistry.ts- Manual images, page numbers, and display rules.src/engines/AIPlanner.ts- High-speed AI intent classification.src/engines/RiskAssessmentEngine.ts- Smart warnings for common high-risk mistakes.
- MIG: Ground clamp -> negative (-), MIG Gun -> positive (+), Gas required.
- Flux-core: Ground clamp -> positive (+), Wire feed -> negative (-), No gas.
- 240V at 200A: 25% duty cycle (2.5m weld / 7.5m rest).
1. Diagrams for wiring.
Polarity is spatial. Users should see which lead goes into which socket. We use React components, not generated SVGs.
2. Short answer first.
The user is in a garage. The first sentence should answer the question. The extra UI carries the detail.
3. No fake precision. When exact wire speed is not grounded in the manual, VO guides the user to the machine's synergic settings instead of hallucinating numbers.
Built for "Gloves-On" utility, VO optimizes for Latency-to-Action.
| Layer | Model Option | Latency | Role |
|---|---|---|---|
| Intent Planning | Gemini Flash-Lite / Groq (Llama) | < 150ms | Sub-100ms command routing & slot extraction. |
| Vision Analysis | Claude Sonnet / Gemini Pro | < 2s | High-fidelity weld bead & hardware audit. |
| Audio Gen | ElevenLabs / Fish Audio | < 3s | Ultra-realistic technical speech synthesis. |
| Topology Rendering | React / Framer | Instant | Deterministic Digital Twin visualization. |
| Voice Buffer | Web Audio API | Parallel | Hands-free shop floor orchestration. |
# Clone and enter repo
git clone https://github.com/ashishsandhu/vulcan-os.git
cd vulcan-os
# Install dependencies (Node.js >= 20.9.0 required)
npm install
# Setup environment
cp .env.example .env
# Launch the engine
npm run devTip
One-Key Power: You only need one API key (Groq or Google) to unlock both the reasoning engine and the voice transcription system.
You only need one LLM provider key to boot the system. Groq and Google are "one-key" solutions that handle both reasoning and transcription:
- Groq (One-Key Ready): Ultra-low latency reasoning + Vision support + built-in Whisper transcription.
- Google / Gemini (One-Key Ready): High-performance Reasoning + Vision support + built-in audio transcription.
- Anthropic (Primary): Best-in-class reasoning and optimized vision (requires Groq/Google/OpenAI for transcription fallback).
- ElevenLabs / Fish Audio: Required only if you want voice synthesis enabled.
Run these test queries to confirm calibration:
- Duty Cycle:
What's the duty cycle for 200A on 240V? - Polarity:
How do I wire for TIG? - Troubleshooting:
I'm seeing pinholes in my weld.
[Screenshot showing the split-pane Chat and Canvas interface]
The app is built as a support cockpit:
- Chat on the left: High-context engineering dialogue.
- Visual workspace on the right: Real-time Digital Twin diagrams and diagnostic cards.
- Inline tables and checklists: Actionable steps inside the conversation.
- Diagnostic Canvas: Persistent view for duty-cycle, polarity, and vision results.
- Image upload: Multimodal entry for live weld diagnosis.
- Voice dictation: Hands-free shop floor utility with sentence buffering.
This is meant to feel like a technical support expert standing next to the welder, not a document search page.
I am an AI Infrastructure Engineer and 2x founder. I treat LLMs as Recursive Thinking Partners to "red-team" constraints and failure states. I don't build wrappers; I build systems that compound in value over time.
- Seniority: Building with generative models since the GPT-2/BART era (2017).
- Scale: Architected systems supporting 10M+ MAU and 25k+ concurrent users.
- Recognition: UK National Digital Startup of the Year Winner & AI Entrepreneur of the Year Finalist.
- Talent Status: O-1A Visa eligible (ensuring frictionless sponsorship for elite talent).
