Kinetic Capacitor (Battery) Guardian: Autonomous System

Inspiration

In the high-stakes world of heavy mining, silence is expensive. When a 36-ton hybrid excavator stops moving, it costs the operation anywhere from $15,000 to $500,000 per hour.

I was inspired by the invisibility of failure in modern renewable energy systems. The Komatsu HB365LC-3 Hybrid Excavator uses a massive Ultra-Capacitor assembly to capture swing energy.1 Unlike a diesel engine that smokes when it breaks, a capacitor fails silently. Its internal chemistry degrades, its Equivalent Series Resistance (ESR) rises, and it enters a state of Thermal Runaway—often exploding before a standard temperature sensor can trigger an alarm.

Current analytics solutions are passive; they show a dashboard after the machine has failed. I wanted to build an immune system for the machine. I wanted to shift the paradigm from Passive Observation to Agentic Resilience, building a system that can see (Vision), hear (Voice), and act (IoT) to save the asset without human latency.

What it does

The Kinetic Capacitor Guardian is an autonomous, multimodal resilience system. It acts as a digital nervous system for hybrid heavy machinery.

Voice-Activated 3D Control: The operator interacts with the system using Agentforce Voice. I can speak naturally—"System status check"—and the 3D Digital Twin responds instantly.

It Sees the Invisible (Physics & Vision): The system runs a Backend Python Simulator that generates real-time telemetry (Voltage, Current, Temperature). Simultaneously, it uses Google Cloud Vision API to analyze camera snapshots from the machine, visually detecting smoke, fire or physical deformation that sensors might miss.

It Visualizes Risk Immersively: Using Frontend HTML Controls, I can inject faults (like "Thermal Runaway"). As the risk score rises, the digital excavator's capacitor bank physically glows red and pulses on the screen, driven by real-time data fetched via Tableau VizQL.

Autonomous Agentic Action (The "Closed Loop"): When the risk hits a critical threshold (TRRS > 0.85), the Agentforce Hybrid Specialist executes a complex, four-step remediation workflow autonomously:

First: It checks Inventory in Data Cloud, confirming a replacement capacitor (Part #CAP-365-HV) is actually in stock.

Second: It creates a Critical Case (Ticket) in the Service Cloud database to officially log the event for service management.

Third: It dispatches a Technician via Salesforce Field Service (SFS), assigning the job to the nearest available worker with high-voltage certification.

Fourth: It simultaneously sends an IoT Command back to the machine, derating the torque to 70% to physically arrest the heat generation.

How I built it

I utilized a Hub & Spoke architecture to decouple heavy processing from the user experience.

The Edge (Python & HTML):

I built a Backend Python Simulator using FastAPI to act as the machine's ECU. It generates the physics data and accepts HTTP commands. The Frontend HTML dashboard allows me to visualize the machine and simulate failures (injecting high resistance) to test the system's reaction.

The Hub (Salesforce Data Cloud):

I used Data Cloud as the central brain. I configured the Ingestion API to receive the JSON telemetry. I wrote Calculated Insights (SQL) to perform the physics math on the streaming data, deriving the risk score:

TRRS = (0.6 x Norm(T_core)) + (0.4 x Norm(R_est))

Spoke 1: The Kinetic Twin (Headless BI):

I rejected the idea of embedding a standard dashboard. I built a Next.js 16 application with React Three Fiber to render the GLTF model of the excavator.

I used the Tableau VizQL Data Service to fetch raw data arrays to drive the animation.

I wrote a custom GLSL Fragment Shader that binds the VizQL data to the mesh's color uniforms.

Result:

Risk → Glow Intensity

.

Spoke 2: The Senses (Vision & Voice):

To make the system multimodal, I integrated Google Cloud Vision to process image snapshots. For interaction, I integrated ElevenLabs to give the Agent a realistic voice, allowing hands-free communication for busy technicians.

Spoke 3: The Agentic Hand (Agentforce & SFS):

I configured an Agentforce Service Agent grounded in the official Komatsu Service Manual using Vector Search (RAG). I connected this agent to the physical world via Apex Actions:

CheckInventory(): Queries Data Cloud objects.

CreateCase(): Inserts a record into the Service Cloud database.

DispatchTech(): Triggers the SFS scheduling logic.

IoT_Derate(): Sends the HTTP POST to the Python simulator.

Challenges I ran into

Multimodal Latency: Synchronizing the Agentforce Voice response, the 3D Visual update, and the IoT Command was difficult. I had to use optimistic UI updates in React to make the interface feel instant while the Data Cloud batch ingestion processed the heavy physics math in the background.

Vision False Positives: The Google Cloud Vision API would sometimes mistake dust for smoke. I solved this by implementing a "Cross-Verification" logic in the Agent: The Agent only triggers a critical shutdown if both the Vision API (Smoke) AND the Telemetry (High Temp) agree.

Trusting the Agent: The biggest challenge was safety. I couldn't let an AI just shut down a massive machine. I configured the Einstein Trust Layer guardrails to strictly block "Shutdown" commands while permitting "Derate" (Limp Mode) commands, ensuring the system failed safely.

Accomplishments that I'm proud of

The "Glow": Seeing the 3D model pulse red in perfect sync with the simulated voltage spike was a "magic moment." It proved that analytics can be visceral and intuitive.

The Full Closed Loop: Watching the logs confirm that the Agent actually checked inventory, created the ticket, dispatched the SFS technician, and derated the machine—all within seconds of the fault injection. It validated the entire "Ingest-to-Action" pipeline.

Giving Analytics a Voice: Hearing the Agent verbally explain why it was derating the machine (citing the Service Manual via ElevenLabs) made the AI feel like a true partner rather than a black box.

What I learned

The Power of Headless: Analytics doesn't have to look like a report. Tableau VizQL Data Service opens up a world where data controls experiences—games, AR, and Digital Twins.

Agentforce is an Orchestrator: It's not just a chatbot. When combined with Salesforce Field Service and Data Cloud Inventory, it becomes an autonomous operations manager.

Physics-Aware AI: Successfully grounding the Agent in the physical reality of the machine (Joule's Law) and visual reality (Computer Vision) created a system that is far more robust than text-only LLMs.

What's next for Kinetic Capacitor (Battery) Guardian: Autonomous System

Edge Integration: Moving the initial computer vision inferencing to the edge (on the machine) using a lightweight model to reduce latency.

AR Overlay: Porting the React Three Fiber application to a mobile AR view, allowing a technician to point an iPad at the excavator and see the "Heatmap" and "Inventory Status" overlaid on the real metal.

Predictive Maintenance 2.0: Replacing the heuristic SQL formula for Risk with a trained Random Forest model in Einstein Studio, allowing the system to predict failures weeks in advance based on subtle vibration anomalies.

Built With

• agentforce
• agentforce-voice
• apex
• einstein-trust-layer
• elevenlabs
• fastapi
• glsl
• google-cloud-vision-api
• next.js
• platform-events
• python
• rag
• react
• react-three-fiber
• rest-api
• salesforce-data-cloud
• salesforce-field-service
• salesforce-flow
• service-cloud
• tableau-pulse
• tableau-vizql-data-service
• tailwind-css
• three.js
• typescript
• vector-search
• webgl