Retrieval Augmented Generation For Industrial Intelligence
In the past week, I’ve had a handful of chats with heads of innovation at big, heavy, indsutrial companies. It seems like they are all facing similar challenges. But allow me a little preface first.
0. AI Agents Empowered by Human Agency
At bld.ai, we are an AI agent-first professional services firm. Our solutions combine powerful AI agents with skilled humans who genuinely care about their work, pay close attention to detail, and ensure everything is done right.
1. Building on Proven Platforms
Instead of building from scratch, our expert teams thoughtfully select established platforms like Samsara for IoT, DataRobot for ML operations, Supabase for real-time databases, and frontends hosted on Vercel. We integrate powerful AI models from OpenAI on Azure, Gemini on Google Cloud, Anthropic on AWS, or open-source alternatives like LLaMA and DeepSeek.
2. Challenges in Heavy Industries
Our approach helps clients in heavy industries—like mining, oil, and gas—manage vast, complex data. These industries often rely on disconnected systems and fragmented workflows, making standardization difficult. Data ownership is unclear, with multiple stakeholders involved, further complicating data management.
3. Understanding RAG: Simplifying Complex Data Access
To tackle these challenges, we use Retrieval-Augmented Generation (RAG), a method that enables AI models to fetch precise, relevant information from large datasets in real-time. For example, when a technician queries maintenance instructions, instead of scanning thousands of pages, RAG automatically pulls out just the necessary information from manuals or logs.
4. Hot Swapping: Efficient Model Routing
Another key technique we apply is “hot swapping.” Think of hot swapping like a traffic cop directing requests. A simple classifier quickly checks each query, decides if it’s about a specific topic (domain-specific) or something general, and instantly directs it to the appropriate AI model. For instance, a query about drilling safety goes to a specialized model, while a general equipment query goes elsewhere. This ensures speed, accuracy, and efficiency without needing massive models for every possible scenario.
5. Continuous Learning Through Human Feedback
We also leverage reinforcement learning with human feedback, meaning every interaction by an internal expert improves the AI. IT professionals and engineers become implicit trainers as they correct or validate results. Over time, the system naturally prioritizes the most reliable and relevant documents, maintaining accuracy even as documentation continuously updates and expands.
6. OSDU: A Foundation for Standardization
An important initiative supporting this is the Open Subsurface Data Universe (OSDU). OSDU provides a standardized, open-source platform for subsurface and operational data, adopted by major energy companies like Shell and BP, and mining giants like BHP and Rio Tinto. Integrating OSDU with ML platforms like DataRobot creates a seamless environment where predictive models deploy rapidly, safely, and effectively.
7. Real-World Application 1: Australia to Singapore Solar Project
Imagine complex projects such as laying an undersea cable connecting solar farms in Australia to power grids in Singapore (see Australia-Asia Power Link). Centralized AI platforms can instantly analyze vast geospatial data, environmental risks, equipment performance data from suppliers like Caterpillar, Komatsu, Sandvik, and logistical bottlenecks. Predictions update continuously in real-time, improving decision-making and reducing project risks.
8. Real-World Application 2: Lula Deepwater Oil Field, Brazil
On the other side of the planet, there’s the Tupi oil field, one of the largest deepwater oil fields off the coast of Brazil and among the most significant offshore discoveries globally. Operated primarily by Petrobras, in partnership with Shell, TotalEnergies, and Galp, and supported by major oilfield services firms including Schlumberger, Baker Hughes, and Halliburton, the Lula field presents complex operational challenges.
Centralized AI platforms integrate enormous volumes of subsea seismic surveys, drilling sensor data, reservoir models, and equipment telemetry. These systems continuously provides real-time insights, predicting optimal drilling locations, proactively identifying equipment maintenance needs, modeling environmental risks, and dynamically updating forecasts. Leveraging this AI-driven intelligence streamlines operations, dramatically reduces downtime, minimizes environmental impact, and substantially lowers project costs, enhancing overall production efficiency.
9. Mars Mindset: Innovation Without Legacy Constraints
At bld.ai, our mindset comes from thinking and talking to experts about how you’d build systems on Mars, free of legacy constraints. This perspective allows us to design simpler, cleaner solutions for Earth’s complex industries. Whether you’re managing subsurface exploration, pipeline maintenance, or predictive safety analytics, our AI systems streamline your data, making it actionable and intelligent.