Inspiration
Most AI systems observe the physical world through dashboards, but they rarely participate in the physical control loop. We wanted to build something different: an AI system that can sense physical infrastructure, understand what is happening, make a safe decision, act, verify the result, and learn from the outcome.
Atlas Sentinel was inspired by the idea of extending the Atlas architecture into Physical AI. Rather than building a generic chatbot or IoT dashboard, we focused on one concrete problem: keeping a small water infrastructure system operational.
Water is an ideal starting point because it is measurable, physically demonstrable, and directly connected to an essential human service. The goal became simple:
“Keep this water system operational.”
From that objective, Atlas Sentinel continuously asks: What is happening? What is abnormal? What might fail? What action is available? Is that action safe? Did it actually work?
What it does
Atlas Sentinel is an autonomous edge intelligence system for resilient physical infrastructure.
Our prototype uses a physical water node with a tank, pump, valves, and sensors for level, flow, and pressure. A microcontroller collects and filters telemetry while an edge computer handles local intelligence and safety-critical logic.
When an anomaly occurs, Atlas Sentinel:
Senses → Filters → Interprets → Reasons → Checks Safety → Acts → Verifies → Learns
The system can detect a controlled failure such as restricted flow or abnormal pump behavior, identify the affected asset, retrieve historical events, reason about likely causes, assess risk, and generate possible interventions.
Only pre-approved low-risk actions can be executed automatically. More consequential operations require human approval.
After an action, the system verifies the new physical state rather than assuming the AI succeeded. The incident, action, outcome, and lesson are then stored so future incidents can benefit from previous experience.
A digital twin mirrors the physical system, while a handheld field interface provides technicians with telemetry, history, evidence, recommended inspections, and incident workflows.
How we built it
We designed Atlas Sentinel as a cyber-physical system with a deliberate edge/cloud split.
The physical layer consists of a water tank, pump, valves, sensors, microcontroller, and actuator controls. The edge layer performs sensor filtering, threshold checks, immediate safety logic, anomaly detection, actuator protection, and offline buffering.
The intelligence layer uses an NVIDIA open-source model such as Nemotron for meaningful reasoning tasks including incident interpretation, diagnosis, historical correlation, action planning, and natural-language explanations. Nebius provides the scalable inference and compute layer through services such as Token Factory or Nebius AI Cloud.
The central Sentinel agent orchestrates the workflow and works with focused components including the Diagnostic Agent, Safety Guardian, and Verifier.
We implemented typed tools around system state, sensor data, history, anomaly detection, diagnosis, risk assessment, safety policies, action execution, verification, and outcome recording.
Most importantly, the AI recommendation is never trusted directly. The flow is:
AI recommendation → Safety validator → Authorization → Local controller → Actuator
This ensures that critical physical safety logic remains local rather than depending entirely on the cloud.
Challenges we ran into
The biggest challenge was making the system genuinely physical rather than simply creating an AI interface around simulated data.
We had to think carefully about the boundary between probabilistic AI reasoning and deterministic control. A language model is useful for interpreting complex situations and planning responses, but it should not be the sole authority controlling hardware.
Another challenge was designing a reliable demonstration. The failure scenario needed to be deterministic enough to reproduce while still showing genuine sensing, reasoning, physical action, and verification.
We also had to decide what belonged on the edge and what belonged in the cloud. Latency-sensitive safety decisions must remain local, while deeper reasoning, historical correlation, simulation, and long-term learning can use scalable cloud infrastructure.
Finally, we wanted every major decision to be explainable through observations, evidence, confidence, unknowns, and action rationale rather than presenting an unexplained AI output.
Accomplishments that we're proud of
We are proud that Atlas Sentinel goes beyond a traditional sensor dashboard or conversational AI demo.
The system demonstrates a complete cyber-physical loop:
Sense → Reason → Act → Verify → Learn
We created a physical water infrastructure prototype where an anomaly can result in a real actuator response, followed by verification of the new physical state.
We also built safety into the architecture from the beginning. The system distinguishes between actions that can be automatically executed and consequential operations that require human approval.
Another accomplishment is the combination of physical telemetry, AI reasoning, a digital twin, evidence, human oversight, and persistent operational memory in one coherent architecture.
Most importantly, the prototype demonstrates the central idea behind Atlas Sentinel:
The AI is not just observing the physical world. It is operating inside it.
What we learned
We learned that Physical AI is not simply about putting an AI model next to a sensor.
The difficult and valuable part is creating a reliable loop between physical state, inference, policy, action, verification, and memory.
We learned that deterministic systems remain essential. AI is strongest when it handles reasoning and ambiguity, while local controllers and explicit safety policies enforce hard constraints.
We also learned that verification is critical. A system should never assume an action worked just because a model predicted that it would. The physical state must change and provide evidence that the intervention had the intended effect.
Finally, we learned that memory can turn repeated incidents into institutional knowledge. When the same anomaly appears again, the system can retrieve what happened previously, what intervention was used, and whether it worked.
What's next for Atlas Sentinel
Our immediate goal is to make the water node more robust, improve the quality of anomaly diagnosis, expand the historical learning system, and strengthen the digital twin and field-operator experience.
Beyond the prototype, the architecture is intentionally designed to generalize beyond water infrastructure.
The same Sentinel architecture could eventually support distributed systems across food, energy, housing, agriculture, manufacturing, and ecological infrastructure.
The near-term vision is not to build every one of those systems at once. It is to make one water node extraordinarily well and prove that the architecture can scale.
The long-term vision is:
Protect essential services by giving physical infrastructure the ability to sense, understand, protect, act, verify, and learn.
Log in or sign up for Devpost to join the conversation.