BatteryForge AI

Inspiration

Battery failures cost industries billions annually and pose serious safety risks, from Samsung's Galaxy Note 7 recalls to Tesla's thermal runaway incidents. Traditional battery quality control relies on manual inspection, destructive testing, and reactive maintenance—all slow, expensive, and error-prone. We envisioned an AI-powered platform that could:

Detect defects visually before they escalate to catastrophic failures
Optimize charging strategies using physics-based simulation to extend battery life
Predict failures proactively using multimodal AI analysis
Coordinate specialist AI agents to handle the complexity of battery manufacturing and fleet management

The breakthrough came from Google's Gemini 3 API & Agent Development Kit (ADK), which enabled building a multi-agent orchestration system where specialized AI "experts" collaborate seamlessly—just like a real battery engineering team.

What it does

BatteryForge AI is a comprehensive battery intelligence platform powered by Gemini 3 and ADK multi-agent orchestration. It provides:

🎯 Core Features

1. AI Commander (ADK Multi-Agent Orchestration)

BatteryForgeCommander: Root agent that routes requests to 5 specialist agents
DefectAnalysisAgent: Visual inspection expert using Gemini Vision
ChargingOptimizationAgent: Electrochemistry and EIS analysis specialist
FleetCommanderAgent: Strategic fleet management and scenario simulation
SafetyGuardianAgent: Emergency response with Human-in-the-Loop (HITL) controls
PredictiveMaintenanceAgent: Lifecycle prediction and maintenance scheduling

2. Visual Intelligence

Real-time thermal runaway detection via webcam, screen share, or YouTube video analysis
"Detect-Locate-Describe" methodology for precise defect classification
PCB/BMS manufacturing defect detection (open circuits, shorts, solder mask issues)
Live streaming analysis with persistent audit trails

3. Charging Analysis & Physics Simulation

PyBaMM integration for Doyle-Fuller-Newman (DFN) physics-based battery simulation
Interactive multi-plot visualization with Recharts (voltage, current, temperature curves)
EIS (Electrochemical Impedance Spectroscopy) analysis with layer-by-layer diagnosis
Universal CSV parser that auto-detects data types (cycling, impedance, telemetry)

4. Fleet Monitor

Real-time monitoring of 100+ battery packs
Scenario simulation (heat waves, cold snaps, fast charging stress tests)
Strategic risk assessment with thermal spread analytics
Pack-level drill-down with maintenance recommendations

5. PCB Manufacturing Automation

Gerber file analysis for CAM validation
Adaptive etching control (conveyor speed, spray pressure optimization)
Lamination scaling prediction for multi-layer boards
Automated compliance certificate generation

6. RAG Knowledge Assistant

ChromaDB vector store with Gemini text embeddings
Ingests PDF manuals and technical documentation
Context-aware responses in chat interface
Semantic search across battery safety standards

How We built it

Architecture

Architecture Diagram

Backend (Python + FastAPI)

ADK Integration: Multi-agent orchestration with Runner.run_async event streaming
Services Layer: Modular services for vision, simulation, RAG, fleet management
Tools: 20+ specialized tools for agents (vision analysis, PyBaMM simulation, EIS processing)
Database: SQLite for analysis history, ChromaDB for vector embeddings

Frontend (React + Vite)

Modern UI: Framer Motion animations, glassmorphism, dark mode
Interactive Charts: Recharts for scientific data visualization
Real-time Features: WebSocket for live agent trace, streaming analysis logs
Components: 10+ specialized components (VisualScout, ChargingAnalysis, FleetMonitor, etc.)

Key Technologies

AI/ML Stack

Google Gemini 3 Flash Preview (multi-agent orchestration)
Google ADK (Agent Development Kit) for workflow coordination
Gemini Vision for multimodal defect detection
ChromaDB + Gemini Embeddings for RAG

Physics & Simulation

PyBaMM (Python Battery Mathematical Modeling) for DFN simulations
NumPy/Pandas for data processing
Custom EIS analysis engine following IEST standards

Frontend Stack

React 18 + Vite (fast development)
Recharts for scientific plotting
Framer Motion for premium animations
React Player for video analysis

Development Process

Phase 1: Built core defect detection using Gemini Vision with "Detect-Locate-Describe" methodology
Phase 2: Integrated PyBaMM for physics-based simulations and universal CSV parsing
Phase 3: Developed Visual Scout with real-time webcam/screen share analysis
Phase 4: Created PCB manufacturing workflows (Gerber analysis, etching control, plating optimization)
Phase 5: Implemented RAG knowledge base with ChromaDB
Phase 6: ADK Multi-Agent System—Migrated to full agentic architecture with 5 specialist agents

Challenges We ran into

1. ADK Tool Schema Generation

Problem: The ADK framework generates JSON schemas from Python function signatures. List parameters without proper type hints caused 400 INVALID_ARGUMENT errors:

GenerateContentRequest.tools[0].function_declarations[3].parameters.properties[video_frames].items: missing field

Solution: Standardized all function signatures to use List[str] from the typing module and ensured docstrings matched the type hints exactly. Updated all 20+ tool functions across the codebase.

2. Event Streaming Architecture

Problem: ADK's Runner.run_async returns an async generator of events, not a simple response object. The initial implementation tried to treat it like a synchronous API call.

Solution: Implemented proper async iteration with event collection:

async for event in self.runner.run_async(...):
    if event.actions.transfer_to_agent:
        # Collect agent transfers
    for fc in event.get_function_calls():
        # Collect tool calls
    if event.content and event.content.parts:
        # Collect streaming response text

3. Model Availability

Problem: Initially used gemini-3-pro-preview which wasn't available in the v1beta API, causing 404 NOT_FOUND errors.

Solution: Discovered gemini-3-flash-preview was available and optimized for tool calling. Migrated all 6 agents to use this model after verification.

4. PyBaMM Threading

Problem: PyBaMM's DFN solver is CPU-intensive and would block the FastAPI event loop, causing the entire backend to freeze during simulations.

Solution: Wrapped all PyBaMM calls in ThreadPoolExecutor with asyncio.run_in_executor() to offload physics calculations to worker threads.

5. Screen Share Video Analysis

Problem: YouTube videos can't be analyzed directly due to CORS. ReactPlayer can display them but can't provide raw frames for canvas capture.

Solution: Implemented getDisplayMedia() screen recording. When a user selects a URL video, the app prompts for screen share permission and analyzes the visible tab instead—a genius workaround for external video analysis!

6. Universal CSV Parsing

Problem: Battery research data comes in countless formats (Arbin cyclers, BioLogic EIS, Tesla telemetry). No single parser could handle them all.

Solution: Built a two-stage system:

Gemini-based semantic mapper: Analyzes headers and sample data to identify column meanings
Adaptive parser: Uses Gemini's mapping to standardize any CSV into canonical format

Accomplishments that we are proud of

🏆 Technical Achievements

First Fully Agentic Battery Platform: Successfully orchestrated 5 specialist AI agents using Google ADK, demonstrating real-world multi-agent coordination beyond simple chatbots.
Physics-Integrated AI: Bridged machine learning with electrochemistry by combining Gemini's reasoning with PyBaMM's physics-based DFN simulations—achieving both accuracy and explainability.
Real-Time Thermal Runaway Detection: Built a production-ready system that analyzes webcam/video feeds at 1 FPS for battery safety monitoring—potentially life-saving technology.
Universal Data Ingestion: Created a Gemini-powered intelligent parser that can understand and process any battery dataset format without manual configuration.
Zero-Shot PCB Defect Detection: Achieved reliable open circuit and short circuit detection using only Gemini Vision, no custom training required.

💡 Innovation Highlights

Screen Share Analysis: Novel approach to analyze YouTube thermal runaway videos by recording the browser tab
Interactive EIS Analysis: Layer-by-layer impedance diagnosis (Ohmic, Kinetics, Diffusion) following industry standards
Agent Trace Visualization: Real-time display of agent transfers and tool calls for transparency
Scenario Simulation: Fleet-level stress testing (heat waves, cold snaps) with physics-based degradation

📊 Scale Achieved

20+ AI Tools across 5 specialist agents
10+ React Components with premium UI/UX
100+ Battery Packs simulated in fleet monitor
6 Analysis Modes: Visual, Charging, Fleet, Logs, Aging, PCB

What We learned

Technical Learnings

Agent Architecture Patterns: Building scalable multi-agent systems requires careful separation of concerns. Each agent needs clear responsibilities and well-defined tool sets.
Event-Driven AI: Moving from request-response to event streaming fundamentally changes how you structure async code. ADK's event model is powerful but requires rethinking traditional API patterns.
Physics-AI Hybrid Systems: The best battery intelligence comes from combining data-driven AI (Gemini) with first-principles physics (PyBaMM). Neither alone is sufficient.
Type System Importance: Python type hints aren't just for IDE autocomplete—they're critical for automatic schema generation in frameworks like ADK.
Real-Time Vision Challenges: Analyzing live video at acceptable frame rates while maintaining accuracy requires careful optimization (canvas rendering, async processing, result caching).

Domain Knowledge

Battery Failure Modes: Deep understanding of lithium-ion degradation mechanisms (SEI growth, lithium plating, thermal runaway propagation)
EIS Interpretation: How to diagnose battery health from impedance spectra using frequency-domain analysis
PyBaMM Ecosystem: The DFN model's power and limitations, parameter sensitivity, computational costs
PCB Manufacturing: CAM workflows, Gerber file structure, etching chemistry, lamination physics

Product Design

Context-Aware AI: The best agent experiences provide workspace context (active tab, current data) automatically
Progressive Disclosure: Start with simple interfaces, reveal complexity on demand (e.g., multi-plot analysis)
Visual Feedback: Real-time status indicators and live traces build trust in AI systems

What's next for BatteryForgeAI

Short Term (1-3 months)

1. Enhanced Agent Capabilities

Add DataAnalystAgent for statistical trend analysis
Implement ComplianceAgent for UL/IEC standard verification
Add SupplyChainAgent for material sourcing optimization

2. Advanced Workflows

Marathon agents for long-running tasks (24-hour pack audits, continuous monitoring loops)
Workflow branching based on severity detection
Automated report generation for compliance