Validation · Monitoring · Evaluation

I find where autonomous systems break.

Evaluations, uncertainty quantification, and failure monitors for safety-critical systems, from gigawatt wind fleets to frontier language models.

Researcher, DTU Wind & Energy Systems · Visiting Researcher, Lawrence Berkeley National Laboratory

Resume GitHub Google Scholar Email
781
citations · h-index 14
27 / 7
LLMs / provider families audited
0.92
AUROC on failure modes unseen in training
+72%
validation coverage via risk-targeted testing
47%
less critical failure risk, deployed in the field
Substrate: Physical Fleets

A decade of safety-critical systems

NREL · CU Boulder · DTU Wind & Energy Systems
Adversarial RL · TORQUE 2026, peer-reviewed
Arms-race training diagram: protagonist and adversary policies iterating against each other

Adversarial robustness via self-play

RL adversaries inject worst-case sensor corruption into fleet control; self-play training recovers from −39% power loss to +7.9% gain with no clean-environment penalty.

View framework →
Validation & uncertainty · informs IEC 61400-60
Safe-operation domain diagram: probably-safe, possibly-safe, and dangerous prediction regions with measurement campaigns

Safe operation envelope prediction

Separates epistemic from aleatoric uncertainty to focus expensive validation campaigns on high-risk conditions — 72% more scenario coverage across 5 simulation platforms.

View validation strategy →
Optimization · in production via TOPFARM

Stochastic gradient descent at fleet scale

Novel SGD algorithm designing layouts for 1,200+ unit fleets, 95% faster than legacy methods; in production in tools used by 30+ companies.

View algorithm →
Risk-aware control · deployed in the field

Control under sensor uncertainty

Uncertainty-aware wake-steering control reduced critical mechanical failure risk 47% with a 0.5% revenue gain — cited 100+ times and referenced in industrial patents.

View control framework →
Substrate: Language Models & Agents

Completed & shipped

Finished work only — in-progress research stays in the lab.
LLM audit · PyPI release
Heat-map of unsafe exception-handler counts across 25 models, 20 seeds, and 3 task difficulties

Silent Killers: error swallowing in LLM-generated code

27 frontier and open-weight models across 7 provider families: 50–100% of generated exception handlers silently suppress errors on scientific-computing tasks. Shipped a static-detection CLI.

View audit →
Agentic benchmark · pre-registered
Performance vs. over-eagerness: open-weight models span 6-72% over-eagerness at comparable accuracy

eager-baker: scope calibration in agents

Isolates scope over-eagerness from task accuracy: models spread 20–72% on over-eagerness against ~6 points on accuracy. Every prompting intervention suppressed capability instead of calibrating behavior.

View benchmark →
Failure prediction · AIAA SciTech, under review
Animation: simulated robot manipulation rollout monitored by the visual safety monitor

Visual safety monitor for robot manipulation

A 33K-parameter monitor predicts vision-language-action policy failure from camera corruption — AUROC 0.92 on corruption types held out of training entirely.

View monitor →
Mechanistic interpretability · BlueDot Impact

A feature no linear probe can read

Reverse-engineered a non-linearly encoded concept in a frozen language model using active subspaces, then built a stranger "pinwheel" phase encoding that defeats linear and quadratic probes, verified causally by rotating the activation plane.

View puzzle solution →
One thesis, two substrates

Turbine fleets and language models fail differently, but the discipline of catching failure is the same: define what "working" means, measure it under conditions nobody planned for, and build the monitor that flags the moment the signal leaves the safe envelope. I spent a decade doing this for physical autonomous systems; I now do it for AI systems too.

Ph.D. Mechanical Engineering, CU Boulder, 2022. Entered AI safety via BlueDot Impact's technical program, 2026. Technical lead, IEA Wind Task 55.