Building the Foundations of Physical AI

I recently joined Hamza Mimi, AWS Solutions Architect, on Episode 10 of the AWS for AI Podcast for a conversation about why context is everything when building intelligent systems. We covered dynamic knowledge graphs, AI agents, physical AI and the concept of agentic twins.

From Raw Data to Operational Intelligence

One of the central ideas in the episode is the data maturity journey every organization goes through: collect data, structure it, connect it, then understand it. Raw data on its own has low value. The real breakthrough happens when you start connecting data across systems, understanding relationships between entities, how they evolve over time, and what events drive what outcomes. That's when you move from data to genuine operational intelligence - the ability to understand what is happening across your operations in real time, trace why decisions were made, and act with confidence.

This is exactly what the Garnet Framework makes possible. Two lines of command give you a full contextual data platform built on a dynamic knowledge graph, a semantic model that captures entities and their relationships while allowing continuous updates in near real-time. Unlike static representations that freeze a snapshot, a dynamic knowledge graph evolves as the real-world systems it represents change, giving you a single source of truth across all your operations.

It doesn't require replacing what you already have. Legacy systems, existing platforms and third-party vendors all become context producers feeding into the same knowledge graph, coexisting and interoperating without disruption. Everything is then accessible through a standardized REST API built on NGSI-LD, with built-in time series for historical analysis, geospatial filtering, and multi-condition pattern triggers that fire when data forms specific combinations across your entities, enabling real-time detection of complex operational scenarios the moment they emerge.

The Semantic Layer: Why It Changes Everything

AI is often framed as a data problem. But the deeper challenge is semantic. Data without meaning is just noise, and the gap between having data and understanding it is where most organizations get stuck.

A semantic layer sits between your raw data and the systems that consume it. It defines what entities are, what their properties mean, how they relate to each other, and how they change over time. It turns data into knowledge. When an AI agent queries your system, it's not just getting numbers or strings. It's getting structured, meaningful context about the real world, context rich enough to support context graphs as a trillion-dollar AI opportunity.

This is what NGSI-LD brings to the table. It's not just a data format. It's a standardized way to describe the world through a property graph model with explicit semantics. You define entities, their types, their relationships, and their temporal history. Every agent consuming this data gets the same grounded, machine-readable understanding regardless of where it runs or who built it.

LLMs and dynamic knowledge graphs work best together. LLMs bring powerful language understanding and reasoning. A dynamic knowledge graph with a semantic layer brings grounded operational truth: which entities exist, how they relate, what state they are in right now, and what history has been captured over time. The NGSI-LD standard even provides a built-in discovery mechanism that lets you quickly explore what types of entities exist in your knowledge graph, making it straightforward to combine both approaches and build agents that reason with real precision.

Temporal Information and Decision Traceability

One capability that really opens up new possibilities here is built-in time series: the dynamic knowledge graph does not just store the current state of your world, it tracks how every entity and relationship evolved over time. Every property change, every relationship update, every event has a timestamp and a history.

This matters for two reasons. First, it enables you to track decision-making. When an AI agent takes an action, you can trace back exactly what context it had, what the state of the system was at that moment, and what inputs drove that output. As we scale to systems with thousands or millions of autonomous agents, this kind of auditability becomes essential. Second, it opens the door to foundation models for time series. When you have a dynamic knowledge graph with clean, structured temporal data across your entities, you can apply FM-based time series models to forecast trends, detect anomalies, and anticipate failures across the full connected context of your operations, not on siloed sensor streams.

AI Agents as Context Consumers

The Garnet Framework decouples context producers (sensors, APIs, legacy systems feeding data into the dynamic knowledge graph) from context consumers (dashboards, analytics, AI agents). This matters a lot in practice. If you swap vendors or add new data sources in a different format, your agents don't need to change because they only consume standardized, semantically-defined data.

There's also a feedback loop dimension. An agent can consume context from the knowledge graph, take an action, write its output back, and trigger another agent with updated context. This is how you build event-driven, multi-agent orchestration on open standards without building brittle point-to-point integrations.

Physical AI: Why the Semantic Layer Becomes Even More Critical

Physical AI means the agentic systems we build today, running inside physical machines: robots, drones, autonomous vehicles. These systems don't just read text. They perceive their environment in real time, reason about it, and act on it. The real-world context they need is spatial, environmental, equipment-aware, and continuously changing. Far more complex than anything we feed today to even the most advanced multimodal LLMs. A semantic layer that describes this physical world precisely becomes essential, not optional.

This is also why you can't train physical AI the same way you train LLMs. You can't send a robot into a manufacturing plant and let it learn on the job. You need digital twins of the physical environment to simulate, generate synthetic data, and validate behavior before deployment. The dynamic knowledge graph is what grounds those simulations in operational reality.

We are still early in this journey and a lot will evolve. What I find interesting is how much the quality of context will shape what these systems can actually do. That is the thread running through this episode, and it is something I am continuing to explore.

You can explore the Garnet Framework in our documentation, read more about the agentic twins concept, or connect with me on LinkedIn or through my website.