RoadStar Sentinel

Tagline

Early warning and counterfactual control for trucking operations.

Inspiration

RoadStar Sentinel grew directly out of my research on PCC — Pressure–Chaos–Control, a framework I have been developing for understanding instability in dynamical systems.

The core idea behind PCC is that adaptive systems often contain three competing forces:

• Pressure, which drives change or pushes the system toward a goal
• Chaos, which introduces variability, disruption, and uncertainty
• Control, which regulates the system and keeps it within useful operating bounds

My broader PCC work explores how the interaction among these forces can generate oscillations, transitions, and instability regimes, and how entropy can be used as a measurable signal of that instability.

While looking at the RoadStar Hackathon, I realized trucking and logistics offered a very natural real-world application.

A fleet is constantly balancing pressure from deadlines, utilization, load commitments, and delivery targets against chaotic disturbances such as traffic, breakdowns, driver shortages, and infrastructure outages. Dispatchers provide the control layer by reassigning freight, activating reserve capacity, delaying loads, or changing plans.

That led to the question behind RoadStar Sentinel:

Can we detect when a logistics network is moving toward an unstable regime before a conventional failure alarm occurs — while the dispatcher still has enough control authority to intervene?

RoadStar Sentinel is my attempt to turn the theoretical PCC idea into a practical logistics decision-support prototype.

What it does

RoadStar Sentinel monitors the evolving state of a trucking network and estimates several quantities:

• Pressure — operational strain caused by deadlines, utilization, and disruptions
• Predictive uncertainty / entropy — uncertainty in the fleet's possible future states
• Control authority — how much useful intervention capacity remains available
• Instability — whether those signals together indicate that the system is entering a dangerous operating regime

Instead of waiting for a hard failure threshold, Sentinel tries to detect deteriorating conditions earlier.

When a warning is triggered, Sentinel evaluates possible dispatcher actions such as:

• reassigning a load,
• using reserve capacity,
• delaying freight,
• rerouting operations,
• or taking no action.

It then runs counterfactual simulations for those choices and ranks them according to how they affect the projected future state.

The dispatcher remains in control.

Sentinel is designed around the workflow:

Detect → Explain → Compare → Override → Replay → Decide

The system therefore acts as a decision-support tool rather than an autonomous dispatcher.

How we built it

RoadStar Sentinel is built as a lightweight Python application with a FastAPI backend and a browser-based operations dashboard.

The core system contains:

• a simplified fleet and freight model,
• a stochastic simulator for future operational states,
• a PCC-inspired pressure/instability engine,
• predictive entropy calculations,
• an early-warning detector,
• a counterfactual action generator,
• a closed-loop control simulator,
• a dispatcher copilot,
• an interactive incident Scenario Lab,
• and an Operations Center visualization.

The Scenario Lab allows the user to inject disturbances such as truck breakdowns, traffic surges, hours-of-service shortages, depot outages, and compound shocks.

Sentinel then simulates how the disruption propagates through the system and shows the evolution of pressure, uncertainty, control authority, instability, and cascade risk.

Candidate dispatcher interventions are evaluated using paired simulation conditions so that actions can be compared more fairly rather than simply receiving different random futures.

The project is intentionally CPU-first and does not require a GPU or cloud-hosted language model for its core functionality.

Challenges we ran into

One of the biggest challenges was defining what an early warning should actually mean.

Detecting a truck breakdown is easy once the truck has already failed. Detecting a missed delivery is easy once the deadline has already been crossed.

The more interesting problem was detecting when the overall system is beginning to lose stability, even though no individual component may yet have reached a catastrophic threshold.

That required combining multiple signals rather than relying on a single alarm.

Another challenge was separating genuine intervention effects from stochastic simulation noise. If two dispatcher actions are evaluated under completely different random futures, one can appear superior simply because it received an easier scenario.

We therefore structured the counterfactual comparisons so candidate actions can be evaluated under matched disturbance conditions.

A third challenge was preserving the human operator's role.

It would have been easy to build a system that simply outputs a single recommended action. Instead, I wanted Sentinel to expose alternatives, explain why they differ, and allow the dispatcher to override the recommendation and see the consequences.

Accomplishments that we're proud of

I am especially proud that RoadStar Sentinel became more than a visualization of PCC.

It now demonstrates a complete closed-loop workflow in which the system can:

• detect developing instability,
• identify the factors contributing to it,
• warn before a modeled conventional failure threshold,
• generate feasible interventions,
• simulate the consequences of each action,
• rank those interventions,
• explain its recommendation,
• allow the dispatcher to override it,
• and replay the resulting future trajectory.

I am also proud that the project connects a theoretical research idea with a concrete operational domain.

My PCC research asks whether instability can be structurally understood, quantitatively measured, and ultimately controlled. Those questions are explicitly part of the research direction of the PCC framework.

RoadStar Sentinel turns those questions into something interactive: a dispatcher can actually watch pressure build, uncertainty increase, control capacity shrink, and then test whether an intervention changes the trajectory.

What we learned

The biggest thing I learned is that optimization and resilience are not the same problem.

During normal operations, the goal may be to minimize travel time, fuel usage, or cost.

During a cascading disruption, the more important question may become:

How much ability to recover does the system still have?

That made control authority one of the most interesting variables in the project.

A system can experience high pressure without failing if it still has enough flexibility to respond. But when pressure rises at the same time that uncertainty increases and available control options disappear, the system may be approaching a much more dangerous regime.

I also learned that counterfactual explanation is extremely useful for human-in-the-loop systems.

Telling a dispatcher "choose Action A" is much less informative than showing:

• what Action A is expected to do,
• what Action B would do,
• what happens if nothing is done,
• and how confident the system is in those comparisons.

Finally, this project gave me a useful new application domain for PCC itself.

The original PCC work has explored biological systems, engineering control systems, and other adaptive systems. RoadStar Sentinel suggests that logistics networks may be another interesting environment for studying Pressure–Chaos–Control dynamics.

What's next for RoadStar Sentinel

The current prototype uses a deliberately simplified, synthetic logistics environment so that the core instability and counterfactual-control ideas can be tested transparently.

The next step is connecting Sentinel to realistic trucking datasets and operational systems.

That could include:

• real orders and trips,
• truck and trailer characteristics,
• driver availability,
• traffic information,
• telematics,
• historical disruptions,
• and dispatch-management systems.

I would also like to extend Sentinel beyond evaluating one action at a time into model-predictive control, where the system evaluates sequences of possible dispatcher decisions over a longer planning horizon.

Other future directions include:

• learning PCC thresholds from historical data,
• richer road-network simulation,
• calibrated uncertainty estimates,
• realistic business costs,
• multi-depot fleet models,
• real-time API integrations,
• and validation against actual logistics disruptions.

More broadly, RoadStar Sentinel is also an experiment in taking the PCC research program from theory into application.

The long-term question is larger than trucking:

Can Pressure–Chaos–Control provide a useful early-warning and decision-support language for complex systems before instability becomes failure?

RoadStar Sentinel is one attempt to find out.

Built With

• ai
• decision
• fastapi
• fleet
• predictive
• predictive-analytics
• python