I design, evaluate, and research AI-enabled systems operating under complex, uncertain, and high-consequence conditions.

My work sits at the intersection of AI systems engineering, model evaluation, adversarial testing, cognitive systems, decision architecture, operational resilience, and human-machine interaction.

I am particularly interested in what happens before obvious failure: when a model, command structure, organization, or coupled human-AI system still appears functional while its underlying ability to interpret, coordinate, adapt, or recover is deteriorating.

Current research spans two connected questions:

How do we know when a system is becoming unstable while meaningful intervention is still possible?

and

How does the architecture of a command system change what humans and machines can still perceive, decide, coordinate, and recover?

Experimental simulation of command-system behavior under degraded, uncertain, adversarial, and dynamically changing operational conditions.

CSIM treats command as a dynamic system rather than measuring effectiveness solely through individual decision correctness or terminal mission outcomes.

The simulation models interactions among:

• information quality
• communications capacity
• uncertainty
• cognitive and organizational load
• authority distribution
• coordination demand
• operational tempo
• system dependencies
• human compensatory capacity
• accumulated degradation
• recovery behavior

A central design principle is the separation of observable performance from latent system state.

A command system can therefore continue producing apparently normal operational output while the underlying architecture is losing resilience, coherence, slack, or recovery capacity.

CSIM currently investigates:

Phantom Stability Observable performance remains stable while underlying resilience deteriorates.

Human Buffer Collapse Human operators temporarily compensate for architectural weakness until finite cognitive and organizational reserve is exhausted.

Coordination Debt Unresolved coordination requirements accumulate and create additional future command workload and recovery burden.

Silent Slack Erosion Adaptive capacity disappears before conventional performance indicators expose degradation.

Command-System Incoherence Locally functional command nodes develop incompatible operational models, producing contradictory action, resource contention, rework, and additional command demand.

Path Dependence Identical disruptions can produce different outcomes because the system reaches them through different prior trajectories.

The simulator allows controlled comparison of command interventions from identical system states, including decisions affecting:

• shared-state synchronization
• immediate execution throughput
• local authority
• command slack
• coordination burden
• recovery potential

Actions are not assigned universal "best decision" scores. Their effects emerge from the interaction among the action, architecture, operational environment, current system state, and previous trajectory.

CSIM also examines how present decisions alter future maneuver through dimensions including:

• reversibility
• recoverability
• resource commitment
• dependency burden
• time to correction
• authority flexibility
• adversary sensitivity
• pathway exclusivity
• constraint accumulation
• ability to reorient
• opportunity creation and destruction

Live Simulator

https://csb1105.github.io/command-simulation/

Source & Research Notebook

https://github.com/csb1105/command-simulation

Interactive stability-analysis environment for examining how pressure, resilience, alignment, divergence, coupling effects, and intervention timing alter complex-system trajectories.

The simulator allows direct manipulation of system conditions to identify dominant destabilizing forces, observe threshold formation, compare interventions, and examine recovery behavior before visible failure.

Live Simulator

https://csb1105.github.io/drift-stability-visualizer/intervention-simulator.html

Source

https://github.com/csb1105/drift-stability-visualizer

Interactive visualization of operational drift, correction capacity, intervention windows, and failure boundaries.

The visualizer explores the period in which a system may remain observably functional even as its capacity to correct or recover is disappearing.

Live Visualizer

https://csb1105.github.io/drift-stability-visualizer/

Source

https://github.com/csb1105/drift-stability-visualizer

My AI evaluation work focuses on the behavior of advanced AI systems under conditions where straightforward benchmark success is insufficient.

Research and evaluation interests include:

• LLM reasoning
• behavioral reliability
• instruction adherence
• adversarial evaluation
• edge-case behavior
• constraint satisfaction
• model failure modes
• factual grounding
• comparative model analysis
• evaluation design
• human-AI interaction
• system-level AI failure

I am particularly interested in distinguishing surface-level output defects from deeper failures involving reasoning, task formulation, context, system architecture, or human-machine interaction.

Structured adversarial evaluation framework for Large Language Models incorporating prompt suites, failure-mode analysis, evaluation artifacts, and AI assurance methodologies.

https://github.com/csb1105/ai-redteam-artifacts

AI evaluation research and experimental artifacts exploring model behavior, reasoning, reliability, and evaluation methodology.

https://github.com/csb1105/evals

Failure is often a late signal.

My pre-failure systems research examines what becomes measurable before a system crosses into obvious or irreversible failure.

The objective is not simply to explain failure retrospectively. It is to identify the changing system conditions that reveal declining resilience while meaningful intervention remains possible.

Research themes include:

• operational drift
• correction capacity
• intervention timing
• recovery landscapes
• state-space modeling
• coupled failure
• threshold formation
• early warning systems
• latent system state
• human compensation
• recovery latency
• path dependence
• decision architecture

Research framework for detecting and modeling instability before failure becomes operationally visible.

https://github.com/csb1105/pre-failure-systems

Framework for modeling destabilizing forces, coupling effects, intervention timing, recovery opportunities, and operational resilience.

https://github.com/csb1105/intervention-stability-system

Risk-aware AI decision-support framework for autonomous mission planning incorporating route generation, threat assessment, constraint validation, mission replanning, and human-AI collaboration.

https://github.com/csb1105/Autonomous-Mission-Planner

Enterprise AI architecture case studies examining implementation strategies, architecture decisions, deployment trade-offs, scalability, governance, and production-system behavior.

https://github.com/csb1105/AI-Systems-Design-Case-Studies

Applied Retrieval-Augmented Generation architecture exploring enterprise knowledge retrieval and AI-assisted information access.

https://github.com/csb1105/enterprise-rag-assistant

AI Research & Evaluation LLM Evaluation • Model Behavior • Reasoning Evaluation • Adversarial Testing • AI Reliability • AI Assurance

AI Systems Engineering AI/ML Systems • Multimodal AI • Autonomous Systems • RAG • Decision Support • System Architecture

Cognitive Systems Cognitive Battlespace • Human-Machine Decision Loops • Meaning Architecture • Interpretive Stability • Cognitive Integrity

Command & Decision Architecture Command Systems • Decision Authority • Future Maneuver • Human Compensation • Coordination Debt • Shared-State Coherence

Complex Systems & Resilience Pre-Failure Systems • Operational Drift • Intervention Timing • Correction Capacity • Recovery • Path Dependence • Coupled Failure

I approach AI as a coupled-systems problem, not simply a model optimization problem.

Models operate inside organizations, missions, interfaces, networks, information environments, authority structures, autonomous components, workflows, and human decision processes.

A technically capable model can therefore participate in an unreliable system.

Likewise, an apparently functional system can already be losing the structural properties required for adaptation and recovery.

For many high-consequence AI applications, the relevant unit of analysis is not the model alone.

It is the coupled human-machine system and the trajectory it is producing.

That changes the engineering question.

Instead of asking only:

Did the model produce the correct output?

I am interested in questions such as:

• What information did the system preserve, distort, or remove?
• What assumptions became embedded in the decision environment?
• Is shared-state coherence increasing or deteriorating?
• What burden is being silently transferred to human operators?
• Which corrective actions remain available?
• Which decisions eliminate future options?
• Is apparent stability being purchased through finite human compensation?
• Can the system still recover?
• What can we observe before failure becomes obvious?

Failure is a late signal.

The more consequential engineering problem is identifying the structural changes that occur while the system is still operating, still adapting, and still recoverable.

Additional work includes:

• Retrieval-Augmented Generation
• Computer Vision
• Predictive Analytics
• Classification
• Time Series Forecasting
• Recommendation Systems
• Customer Intelligence
• Business Analytics
• Machine Learning Model Development

Additional applied AI projects are available through Kaggle.

I publish technical and analytical work on:

• AI systems engineering
• AI evaluation and red teaming
• cognitive systems
• human-machine command
• decision architecture
• operational resilience
• pre-failure behavior
• AI assurance
• autonomous systems
• mathematical models of command

Substack https://genxtechwriter.substack.com

Website https://www.carolinebrooks.org

LinkedIn https://www.linkedin.com/in/csb1105

GitHub https://github.com/csb1105

Kaggle https://www.kaggle.com/carolinesbrooks

Substack https://genxtechwriter.substack.com