Research
Twelve areas.
Real problems.
Machine learning, agentic AI and AI safety first, applied to hard problems in energy, infrastructure and government. Each area below is a set of problems we are actively solving, with the methods we use and where you can see it running.
AI SafetyAgentic AIMachine LearningReinforcement LearningAdvanced MethodsAI DeploymentAI EconomicsGovernmentForecastingEnergy IndustryBuildingsClean-Tech
AI Safety & Governance
Making sure AI can be trusted, audited and governed when the stakes are real.
Problems we work on
- Training AI auditors that uncover hidden behaviors in other models
- Can an AI reviewer catch errors that look internally consistent?
- Validation, sign-off and governance before an AI decision goes live
AI auditingAdversarial evaluationCalibration trainingModel risk validation
Agentic AI
Agents that plan, use tools, act on real data and check their own work.
Problems we work on
- Self-correcting research agents that run for hours unattended
- Decision agents that know how likely they are to be wrong
- Multi-agent pipelines for market and business intelligence
LLM agentsTool useSelf-reflection loopsMulti-agent systems
Machine Learning & Model Design
Choosing and building the right model for the problem, then proving it.
Problems we work on
- Frontier LLM vs. small specialized model vs. classical ML: which wins
- Benchmarks on real data instead of toy sets
- Custom RAG over private documents, with memory and sources
Neural networksGradient boostingBayesian inferenceRAGBenchmarking
Reinforcement Learning & Compute
RL that controls real hardware, and the economics of AI infrastructure.
Problems we work on
- Reinforcement learning that controls GPU power draw
- On-site power vs. the grid for AI data centers
- Tracking data center buildout and power demand
Reinforcement learning (GRPO)Fleet simulationTechno-economics
Advanced AI & ML Methods
State-of-the-art methods for problems where standard models are not enough.
Problems we work on
- Uncertainty-aware forecasting with Bayesian models and Gaussian processes
- Reinforcement learning for control and sequential decisions
- Neural networks and specialized models trained on domain data
Gaussian processesBayesian inferenceDeep learningRL (GRPO)Calibration
AI Deployment & Infrastructure
Helping organizations deploy AI the right way, at the right cost.
Problems we work on
- Claude, GPT or other APIs vs. open-weight models you host yourself
- Sizing GPUs, memory and hosting for the model you choose
- Full deployment cost compared with LCOAI before you commit
Model selectionGPU & RAM sizingInference benchmarkingLCOAI
AI Economics
What AI really costs, and whether it pays back.
Problems we work on
- The full lifetime cost of an AI system on one number
- Build vs. buy vs. API, compared on equal terms
- Whether AI can improve a specific project, and how ready it is
Lifecycle cost modelingBayesian uncertaintyScenario analysis
AI for Government
Safe, accountable and cost-accounted AI for public agencies.
Problems we work on
- A safe gateway between staff and AI models
- Safety costs counted inside the AI budget
- Governance, policy and procurement guidance
Secure AI architectureRisk assessmentSafety-inclusive LCOAI
Forecasting & Energy Markets
Machine learning on prices, fundamentals, projects and market events.
Problems we work on
- Which grid battery projects in the queue actually get built
- Natural gas and power price forecasting (ERCOT spreads)
- Turning market events into structured, usable signals
ML forecastingBayesian modelsCalibrated agentsEvent-driven pipelines
Energy Industry
AI and machine learning for energy companies, utilities, developers and funders.
Problems we work on
- Techno-economics for hydrogen, storage, power and alternative fuels, with full uncertainty
- Screening projects and investments for developers and funders
- Market and policy scenarios across U.S. regions
Bayesian learning curvesQuantile gradient boostingReinforcement learningScenario analysis
Buildings & Efficiency
AI that recommends the right retrofit for a real building.
Problems we work on
- Recommending energy conservation measures from audit data
- Supervised models vs. LLMs on 10,000+ real NYC audits
- Prioritizing retrofits by cost and savings
Gradient boostingRetrieval-augmented LLMsBenchmark design
Clean-Tech Evaluation
Judging new technologies against the market, not just readiness level.
Problems we work on
- Benchmarking emerging technologies beyond TRL
- Comparing technology cost and performance to incumbents
- Due diligence support for funders and investors
ML scoringTechno-economicsMarket benchmarking
Selected publications
Some of it is published.
A few of the papers and reports behind the areas above.
Study 2026Which AI should pick grid batteries? Interconnection-queue forecastingZenodoRead →
FrameworkLCOAI: Levelized Cost of AIarXiv & ElsevierFramework →
arXiv 2026Beyond the Grid: on-site power for AI data centersarXiv:2608.08170Read →
ReportReinforcement learning power control for GPU clustersTR-2026-01Demo →
PreprintThe Consistency TrapUnder review
Working paperRetrieval over Reasoning: LLMs for energy retrofit recommendationLL87 benchmark
JournalHydrogen techno-economics across U.S. marketsInt. J. Hydrogen Energy, 2025
Have a problem in one of these areas?
We take on joint research, sponsored studies and builds with companies, utilities and public agencies.